Levering up on compute

NVIDIA moved three times in one week: a $5B strategic investment in Safe Superintelligence (SSI), $1B for a 4.5% stake in Korea's Naver to build a national AI factory, and a batch of seed rounds. Nearly 90 AI rounds were confirmed and over $14B disclosed this week. NVIDIA is turning from the picks-and-shovels vendor into the single most active investor in the primary market.

Locking down AI agents

AI security was the densest theme, and its target moved. Of 15 AI-security rounds this week, 11 went to locking down non-human identity: AI agents, service accounts, machine identities. The consolidation was just as fast: Cyera agreed to acquire Oasis Security for ~$1B ($700M cash plus stock), and Okta bought Permiso for ~$200M. Security leaders are buying, rather than building, the infrastructure to govern proliferating AI agents.

Money sank into the hardest layers

Commonwealth Fusion Systems added another $1B (to $4B total) to power AI data centers, the largest single round in fusion history. Multiverse Computing raised $570M for quantum model compression, Antora Energy $550M for industrial heat storage, K2 Space $500M for satellite platforms. Capital is settling into energy, compute, and frontier models, the three hardest layers of the stack.

Deleveraging to pay debt

On the way out, the story reversed. The "Situational Awareness" AI hedge fund blew up and deleveraged in public markets; MiniMax halved on its listing debut. Yet Chinese AI assets kept queuing to go public together: Moonshot AI (Kimi, ~$35B post-money) and GPU maker MetaX are both pushing toward Hong Kong listings. The valuations built in private now have to clear public markets, and the public market is cooling first.

Our read

The primary and secondary markets are diverging. The primary market keeps levering up with NVIDIA's strategic capital, fusion's mega-rounds, and dense security bets, wagering on the infrastructure that scaled AI agents will require. The secondary market is paying down two years of valuation excess, and the assets that listed first are hit first. As the tide goes out in public markets, capital concentrates into the top names and the hardest infrastructure. In a deleveraging cycle, leaders become the last safe harbor.

Full data and capital map: https://svtr.ai/funding?utm_source=substack&utm_medium=email&utm_campaign=weekly-168&utm_content=article-card

Levering up on compute

NVIDIA moved three times in one week: a $5B strategic investment in Safe Superintelligence (SSI), $1B for a 4.5% stake in Korea’s Naver to build a national AI factory, and a batch of seed rounds. Nearly 90 AI rounds were confirmed and over $14B disclosed this week (SVTR AI Venture Database #168, https://svtr.ai/funding?utm_source=substack&utm_medium=email&utm_campaign=weekly-168&utm_content=article-card). NVIDIA is turning from the picks-and-shovels vendor into the single most active investor in the primary market.

Locking down AI agents

AI security was the densest theme, and its target moved. Of 15 AI-security rounds this week, 11 went to locking down non-human identity: AI agents, service accounts, machine identities. The consolidation was just as fast: Cyera agreed to acquire Oasis Security for ~$1B ($700M cash plus stock), and Okta bought Permiso for ~$200M. Security leaders are buying, rather than building, the infrastructure to govern proliferating AI agents.

Money sank into the hardest layers

Commonwealth Fusion Systems added another $1B (to $4B total) to power AI data centers, the largest single round in fusion history. Multiverse Computing raised $570M for quantum model compression, Antora Energy $550M for industrial heat storage, K2 Space $500M for satellite platforms. Capital is settling into energy, compute, and frontier models, the three hardest layers of the stack.

Deleveraging to pay debt

On the way out, the story reversed. The “Situational Awareness” AI hedge fund blew up and deleveraged in public markets; MiniMax halved on its listing debut. Yet Chinese AI assets kept queuing to go public together: Moonshot AI (Kimi, ~$35B post-money) and GPU maker MetaX are both pushing toward Hong Kong listings. The valuations built in private now have to clear public markets, and the public market is cooling first.

Our read

The primary and secondary markets are diverging. The primary market keeps levering up with NVIDIA’s strategic capital, fusion’s mega-rounds, and dense security bets, wagering on the infrastructure that scaled AI agents will require. The secondary market is paying down two years of valuation excess, and the assets that listed first are hit first. As the tide goes out in public markets, capital concentrates into the top names and the hardest infrastructure. In a deleveraging cycle, leaders become the last safe harbor.

Full data and capital map.

The real risk is distribution, not total

Even if AI ends up creating far more than $3T in value, the money splits four ways:

• Spending it: a handful of hyperscalers, data centers, frontier labs
• Compressing it: open-weight models, inference optimization, cheaper silicon
• Enjoying the productivity: enterprises and consumers using AI
• Actually earning it: apps that own the user, the workflow, the transaction, and proprietary data

Here’s the view from our own data. Across the 5,718 AI companies in the SVTR AI VC Database, the model and infrastructure layers are only ~22% of firms but hold ~67% of disclosed funding — model-layer firms average ~$1B raised each, about 13x the application layer. Capital sits upstream; revenue has to come from the ~80% of firms downstream.

Efficiency and open weights change who gets paid

The “token factory” assumes usage grows to fill capacity. Two forces in 2026 move the revenue, not the volume.

Efficiency: OpenAI’s GPT-5.6 Sol is 54% more token-efficient on agentic coding than rival models (Altman, July 9). Good for users; for token-metered vendors, the same job may bring in less — unless usage scales up to compensate.

Open weights: on OpenRouter, Chinese models overtook US models in the week of Feb 9–15 and by June ran ~18T vs ~5.5T weekly tokens; the US share fell from ~70% to ~30%. They’re 10–20x cheaper.

Jevons’ paradox all but guarantees total usage keeps exploding. The open question is who gets paid: when the price of intelligence falls faster than demand rises, incremental usage leaks toward models that don’t pay for that $750B of capex.

China isn’t answering the question — it’s making it harder

China runs the opposite playbook: capex-light, open-weight-heavy, priced to win developers. But a nuance that gets lost: the weights are Chinese, the calls often aren’t Chinese-hosted — they still run on AWS, Azure, Nvidia. What open weights compress first is model-layer margin and high-margin API revenue, not necessarily all of infrastructure revenue. Value is moving from model ownership to compute hosting, distribution, and workflow.

Why second-order investors should care

Torsten Slok’s worry isn’t AI itself — it’s transmission. Consensus has the four hyperscalers’ combined free cash flow growing 4x+ from 2026 to 2030. But with the Magnificent 7 so heavy in the index, a slower payoff wouldn’t stay contained — it spreads to chips, power, data centers, and the S&P 500.

The bottom line: AI will likely create far more than $3T in value, but the people who put up the $3T in capital risk won’t necessarily capture a proportional share of the revenue. For the next 12 months, the bet isn’t whose model is best — it’s who can own a high-frequency workflow, accumulate data a model can’t copy, and hold the customer relationship. When tokens are cheap, the call isn’t the moat; embedding it in a business outcome is.

Read the full piece with the AI VC Database charts →

1. The target isn’t the chip, it’s the CUDA dependency

The most quotable line is “Nvidia is digging its own grave,” easily misread as “domestic chips will catch up soon.” Liang’s point is narrower: Nvidia’s strongest software moat is eroding as AI writes code and higher-level languages lower the cost of building and porting kernels. DeepSeek open-sourced TileKernels, built on its TileLang DSL. But TileKernels still depends on Nvidia GPUs and CUDA (13.1+, Hopper and Blackwell). What it erodes is the dependence on hand-written CUDA, not the full stack of drivers, compilers and comms libraries.

So the real claim is that CUDA’s accumulated migration cost may fall. Whether domestic silicon catches the opening still hinges on capacity, real training runs, and migration cost. SVTR’s read: the valuation anchor for domestic compute should move from “can it be adapted” to “can it train.” The only re-rating trigger is a flagship model trained end-to-end on domestic chips, with public results and a reproducible path.

2. Open source is a cost strategy, not idealism

When official serving cost is low enough, releasing weights doesn’t automatically kill the API business. Enterprises compare the total cost of self-hosting against the API price; whether weights are free is secondary. That is why DeepSeek keeps pushing compute efficiency: for compute-constrained teams, efficiency decides how large a model you can train at all. DeepSeek V4 Flash and Pro output tokens run $0.28 and $0.87 per million, under a tenth of comparable OpenAI models.

The uncomfortable read for AI-application investors: the model API layer is deliberately giving up fat margins. If model capability keeps rising and prices keep falling, what does a middle-layer company own that the base model won’t absorb? Reselling tokens or wrapping workflows won’t hold; the durable ones own a data loop, execution, or outcome delivery.

3. The gap isn’t only compute, but compute buys trial-and-error

Liang attributes most of the China-US frontier gap to compute: roughly two years behind, on a twentieth of the compute. Compute doesn’t just size the final run; it decides how many experiments a team clears before training, and how fast it recovers from dead ends. Yet he concedes high-quality data and post-training are the other binding constraint, and capex can’t close those overnight. SVTR’s read: the competitive standard is shifting from “who raised more” to “who converts compute into capability.” Buying cards is only an asset when the cards run and keep producing model progress.

4. After agents, the real step is continual learning

Liang frames progress as a ladder: language models, chain-of-thought, agents, and next, continual learning, still unsolved globally. Today’s agents lean on humans for context; most systems don’t turn a finished task into lasting capability. DeepSeek’s tell is that it builds each model first to accelerate its own R&D, so every generation raises the next one’s development speed. That is closer to what frontier labs are actually racing for than one more agent app. The restraint: continual learning isn’t a “long-term memory” button or a bigger parameter count.

5. One logic of restraint

The four answers differ; the logic is one. DeepSeek keeps base models, efficiency and AGI research inside, hands more application opportunity to partners, uses open source to lower friction and low prices to widen use, and caps commercialization so it never overrides research priority.

It isn’t automatically right. But the scarcity is that Liang explains every choice with one goal instead of packaging open source, low prices, domestic chips and AGI as separate slogans. It shows up in the data too: per SVTR AI Venture Database, DeepSeek’s June A-round (about $56.3B valuation) is led not by financial VCs but by the National AI Industry Investment Fund alongside Tencent, JD, NetEase and battery maker CATL, matching the transcript’s line about picking the most aligned investors. On the talent side, the DeepSeek capital graph and Liang Wenfeng relationship graph show a core team drawn almost entirely from his quant fund High-Flyer and Zhejiang University, the structural basis for “team stability is the only core interest.”

Full analysis: Liang Wenfeng’s Closed-door Talk: NVIDIA Is Digging Its Own Grave

*Seed rounds get bought by strategics.* Walden has had general-purpose robots working in a Toyota plant since February. Who signs the first big check for a company like that? Not Sequoia, but the Toyota that will use it, the CoreWeave that gives it compute, the Prologis that will deploy it. For a company building physical robots, a real factory, reserved compute, and a deployment pipeline are worth more than money, and money is not the scarce thing right now.

*Even stars still raising money have started writing checks upstream.* In China the pattern is more extreme. This week AgiBot, itself still venture-funded and nowhere near an IPO, showed up as an industrial investor in two rounds: Xinqi Robotics (humanoid joint modules) and MoSense (full-body tactile sensing). Both are its own upstream components. The same week Meituan led a RMB 500M round into dexterous-hand maker Xynova. Once being a patron no longer requires profitability or a listing, a Series C company can be the next layer’s bank.

*Financial VCs are losing pricing power.* When a round is led by a strategic that wants to use, supply, or bind you, the valuation carries a “strategic premium” and the exit leans on that strategic absorbing you rather than an independent path to public markets. Independent funds get pushed to three positions: early scout, application-layer software the strategics avoid, or co-lead with ceded control.

The US–China cut is the part other platforms miss. US patrons are mostly mature giants (NVIDIA, Databricks, AMD) plus one neocloud newcomer, CoreWeave. China is supply-chain champions (CATL, Meituan, Shanghai Electric) plus venture-stage stars acting as GPs (AgiBot, SenseTime). One is led by profitable incumbents; the other looks more like picking sides.

Why now: value has concentrated upstream past a threshold, capacity and compute are scarce so non-cash resources are at their most valuable, the exit gate just cracked open (see our earlier piece on China’s capital cycle closing its first loop), and robotics has entered a supply-chain phase where components must be locked in early.

Our read: capital is starting to circulate inside the AI industry, and the independent financial VC is being squeezed out of the lead seat on the best early deals.

Read the full piece with the panorama and three metrics we are tracking

The exit door: the IPO queue got real

Anthropic confidentially filed its S-1 in June, with Morgan Stanley, Goldman Sachs and JPMorgan mandated and bankers already arranging investor meetings, targeting an October listing at a valuation approaching $1T. Data-center operator Switch hired Goldman and JPMorgan for an IPO of up to $10B, as soon as Q4. DeepSeek is reportedly in talks to raise ~$1.5B at ~$71B while preparing an A-share filing as early as year-end. And humanoid-robot maker LimX Dynamics closed a ~$200M pre-IPO round at roughly $2.1B, reportedly heading for Hong Kong.

The signal is one and the same: pricing power over the best AI assets is moving from the private market to public markets. The trillion-dollar valuations stacked up in private rounds now have a date with public-market due diligence, and it moved from rumor to bank mandates this week.

The entry door: strategics packed the lead seats

Walden Robotics came out of stealth with a $300M seed at a $1.1B valuation. The cap table reads like a supply chain, not a fund list: Toyota, NVIDIA, Boeing, Samsung, CoreWeave, Prologis. The company spun out of Toyota Research Institute in January and is led by MIT professor Russ Tedrake.

Of the 84 rounds SVTR tracked this week, about 10 were led directly by strategic or industrial investors, and roughly 20 had industrial capital somewhere on the list. In China it goes further: AgiBot, itself still pre-IPO, invested in two of its own component suppliers (joint modules and tactile sensing); Meituan led a $69M round in a dexterous-hand maker; CATL and Shanghai Electric led robotics rounds. The winners of the last layer are using equity to lock up the next one: supply chains, deployment scenarios, compute quotas.

Where that leaves financial VCs

They still wrote the week’s biggest confirmed check, but note where: Fireworks AI’s $1.5B Series D at $17.5B sits in inference, on a $1B revenue run-rate growing 5x year over year, with NVIDIA merely following. That is the layer strategics have not claimed. And the new powder is going earlier: Greylock closed Fund 18 at $1.5B for seed and Series A; AlleyCorp closed $335M for pre-seed.

Strategics take the lead seats. Public markets take the exits. Financial VCs retreat to day one.

The full issue covers all 72 confirmed rounds (~$3.9B disclosed), the M&A moves (Hyundai buying out SoftBank’s remaining Boston Dynamics stake among them) and this week’s community signals: read the full weekly on SVTR, and explore the live funding feed with charts, updated daily.

The money went physical

The biggest confirmed rounds this week weren’t foundation models or SaaS. They were drones, humanoid robots, satellites and quantum computers.

SVTR’s AI Venture Database logged 86 confirmed rounds (59 disclosed, $7.33B total). Eight of the ten largest are hardware:

• Quantum Systems — $1.2B, military drones (Blackstone & Airbus co-leading)
• Together AI — $800M, AI inference cloud (Aramco, Nvidia, General Catalyst)
• Astribot (智平方) — $735M, humanoid robots
• Chang Guang Satellite — $694M
• Etched — $500M, transformer-only ASIC (Peter Thiel, Jane Street)
• IQM — $234M IPO, quantum computing

Only Together AI and one enterprise-messaging startup are pure software. Robotics, quantum, optics and space fill the rest of the board.

The exits went physical too

Five AI companies IPO’d this week, and 7 of 8 listing / pre-IPO events came from China. Unitree cleared its STAR Market registration in 73 days — the first humanoid / quadruped robot maker to do so on China’s main board. Kuaishou’s Kling closed a ~$3B pre-IPO at a ~$18B post.

This is the closing of a loop. For the first time, China’s AI venture cycle has a working exit anchor, and the scarce resource is shifting from money to a listing path. (We unpack this in this week’s flagship Signal.)

Many-and-small vs few-and-large

A tempting but misleading read is “the top backers are all Chinese.” It’s only half true.

By deal count, Chinese early-stage firms lead — Hillhouse and Sequoia China each backed three rounds. By check size, it flips: the week’s largest rounds were led by Blackstone, Wellington, Nvidia, General Catalyst and Radical. Chinese capital is running a high-volume early-stage factory in hardware; overseas capital is writing a few very large checks. Two different games, same sector.

Why it matters

The model layer is no longer the thing being funded, or the thing being exited. On both the entry and the exit side, capital’s center of gravity moved to things you can touch — robots, chips, satellites, drones.

Full data and the four capital-map charts

The real signal in AI infrastructure this week isn’t another Nvidia challenger. It’s that the largest infrastructure bets all walked around Nvidia’s strongest front.

The two flanks. Qualcomm confirmed a ~$3.9B all-stock acquisition of Modular — a 2022 company that makes no chips. Its product is hardware-agnostic inference software (the Mojo language plus the MAX engine) that lets developers write AI code once and run it across CPUs, GPUs and NPUs without rewriting per chip. A chipmaker that can’t beat Nvidia on raw compute density can still buy the layer that unbinds code from Nvidia. CUDA lock-in is Nvidia’s deepest moat; whoever pries it open makes their own silicon selectable again.

The second flank is light. When single-chip compute is no longer the only scarce thing, how thousands of chips talk to each other becomes the new wall. SpaceX bought one-year-old Mesh Optical (founded by three ex-SpaceX Starlink laser-comms engineers; 1.6Tbps transceivers). MediaTek led HyperLight’s $80M Series C (thin-film lithium niobate photonics). Upscale AI raised a $190M Series A extension for open-standard AI networking — with Nvidia’s own venture arm on the list. None of these fight compute density; they sell connecting, not computing.

Why now. The load shape changed. Training is a peak load that rewards single-chip density — Nvidia’s home turf. Inference is a sustained load that rewards unit cost and the ability to wire huge fleets together efficiently. As the industry shifts from “train a bigger model” to “run models cheaply at scale,” the buyer’s pain moves from “can’t get chips” to two new things: the cost of being locked into one ecosystem, and congestion between chips. When GPU supply was tightest, no one had time for decoupling; as supply eases and inference scales, “don’t get locked in” and “connect the cards well” become fundable businesses.

The China mirror. Where US capital goes around Nvidia, Chinese capital backfills the stack — and the main stage is the secondary market. Enflame won approval for a Shanghai STAR Market IPO; SG Micro (analog) and SinoMicro (lithography gear) completed HK A+H listings on the same day, the latter up 104% on debut. One flanks via M&A and private rounds; the other backfills via public-market funding, one link at a time.

The supply gap. Of the 1,282 founders who’ve applied to SVTR, only ~7% (among those who named a sector) work on compute, chips or semis — the rest cluster in applications. Infra isn’t a founder gold rush; it’s a seller’s market where a few hard teams get chased by capital and acquirers.

Track over the next 90 days: funding and M&A in the decouple-and-runtime layer; quarterly optical-interconnect and silicon-photonics funding; and the share of domestic chip and semi-equipment names in China’s STAR and HK A+H pipeline.

Source: SVTR AI Venture Database, Issue #163 (112 deals).

Everyone is pricing the AI model layer. Goodwater spent a decade pricing the opposite bet.

Goodwater Capital calls itself the world's largest pure consumer-tech VC — ~$3.3B AUM, 79 companies across 19 countries, ~60% in seed and early stage. Its thesis, in co-founder Chi-Hua Chien's words: the biggest winners of the AI era won't sell AI. So it deliberately skips models and compute, and backs consumer applications only.

The historical logic. Chien repeats one rule to LPs: value accrues to the application layer, not infrastructure. Internet era — infrastructure created ~$400B in market cap, applications ~$3.1T. Mobile — infrastructure ~$700B, applications ~$3.7T. Netflix, Meta and Uber all grew on pipes someone else laid.

That distribution is showing up in our own data. Per SVTR's AI venture database, among the AI companies we've tiered, the application layer holds 4,261 companies vs. 571 at the model layer and 671 at infrastructure (as of June 2026). Capital's volume concentrates on that ~10% of model companies; the density of companies actually being built sits in applications. Goodwater is betting that AI won't flip this.

The AI twist: capital efficiency. In 2023 Goodwater raised over $1B in one shot, ~60% for seed and early. The assumption: AI lets application companies hit revenue and margins with tiny teams, so the unit economics of an early check are better than in the last consumer cycle. Chien also leaves infrastructure investors a sharper warning — the capability gap between on-device AI and cloud frontier models was 18–24 months two years ago, is ~6 months now, and he expects it to compress to 3 within a year. Once it closes, the model layer's differentiation commoditizes faster than most expect.

The portfolio: from giant exits to a "personalization" bet. Goodwater's past exits were mostly absorbed by incumbents: Musical.ly (acquired by ByteDance for >$1B in 2017, folded into TikTok), Photomath (Google), Dosh (Cardlytics), Care/of (Bayer). Still-held names span the globe — Monzo in the UK, Toss Bank in Korea. This cycle's AI bets are the part to watch: 22 deals since 2023, almost all seed and early, across entertainment, office, tax, healthcare and finance. Read AI raised twice in 2024 ($21M A, then $50M B six months later); the largest is Stash ($146M H round, 2025). Chien calls the theme "personalization" — entertainment apps doing $100M–$600M ARR at high margins where users don't feel they're using AI; MIDI Health using AI to widen the supply of doctors for perimenopausal hormone therapy. His most contrarian line is "offline reconnection": he backed Fever (a "European Live Nation"), betting that the more infinite digital content becomes, the scarcer and more valuable real-world gatherings get.

The quiet tell. 8 of those 22 deals were co-invested alongside Y Combinator in the same round; Lightspeed and Madrona twice each. So nearly a third of its early bets sit downstream of YC's already-filtered consumer funnel — its real edge is likely less "unique discovery" than judging which app's user retention can survive the giants, and daring to bet early in emerging markets.

The risks — all on one assumption. Consumer-AI moats are thin and get thinner as model capability spills over; differentiation falls back on distribution and brand, the incumbents' home turf. Exits depend on overseas M&A and IPO windows (Getir kept losing money; Monzo had turbulent years), and emerging-market liquidity is more cycle-dependent. And the women's-health / voice / offline-reconnection directions are early signals, not yet proven markets big enough to return a large fund.

Why it matters for cross-border investors. Betting on the application layer is nearly consensus. The hard part is whether the winners are startups at all. When capability is everyone's, the game returns to distribution and user relationships, where Meta and Google already sit. The next 18 months are the test: watch which of Goodwater's bets the giants absorb, and you'll know how deep its moat really is. In China the question barely breathes. Consumer innovation is internalized by ByteDance, Tencent and Alibaba, yet Goodwater is already entangled with Chinese capital (Musical.ly went to ByteDance; portfolio name Weee! serves North American Chinese). Read it as an observation post: watch which country its next seed check lands in.

Sources: SVTR AI venture database (Goodwater record, 22 AI deals, AI company tiering) · Goodwater 2023 fundraise & site · StrictlyVC / TechCrunch interviews with Chi-Hua Chien (June 2026).

Acquirers bought agents, public markets bought scale, and early-stage money bought paradigms with no product yet. The one player caught in the middle — the pure-model company living on narrative premium — is the one getting squeezed.

Buyers paid for agents, not models

Five acquisitions in one week, and not one bought a model. SpaceX took Cursor for ~$60B in all-stock, folding it into the Colossus supercomputer. Salesforce bought Fin (formerly Intercom) for $3.6B to bolt onto Agentforce. Databricks (Panther Labs), 1Password (Apono) and Elastic (DeductiveAI) each bought security or agent capabilities to slot into their platforms. The pricing power has moved: not “whose model is strongest,” but “whose agent already does the work.”

Public markets paid for scale — China edition

The same week, two Chinese AI assets cleared the gate. Xiaohongshu filed in Hong Kong. Enflame, the last of China’s “four AI-chip dragons,” won its STAR Market IPO. Neither is a pure-model company; both are revenue and strategic-position stories. Public markets are pricing cash flow and positioning, not model narrative.

Early stage paid for paradigms

General Intuition raised a ~$300M Series A (reportedly backed by Bezos and Schmidt); CuspAI raised ~$400M for materials-science AI. These are world-model and science bets — funded before there is a product to sell. Capital’s appetite for “a paradigm without a product yet” is unusually high right now.

The middle: OpenAI

OpenAI reported a $38.5B net loss for 2025. Read it carefully: it includes a one-off $41.6B non-cash charge tied to its for-profit transition — strip that out and operations roughly break even. The sharper signal is share: ChatGPT slipped below 50% of the global market for the first time (46.4%), even as it still leads on absolute usage (1.1B MAU).

The honest counterpoint

The single largest confirmed raise of the week was a model company: DeepSeek, at $7.4B — about 74% of all confirmed AI funding we tracked (51 deals, ~$10.05B; 56 companies including 5 rumored rounds). The model layer is not dead. It is bifurcating along US/China lines and concentrating into single points. Geographically the week was a US (23) / China (18) dual core; by count, the application layer dominated (33 companies), while the model layer’s dollars rode almost entirely on DeepSeek.

The squeeze isn’t on “models” broadly. It’s on the middle of the model layer — pure-model companies living on a narrative premium — now pressed from three sides at once: agents above, scale below, paradigm bets ahead.

Capital is pricing the paradigm, not the product — and the US and China are buying two different positions with two different kinds of money.

In SVTR’s AI Venture Database this week (Issue #162, ~56 AI rounds), about ten early-stage deals went to the physical world: world models, embodied AI, robotics, AI materials. Most have little to no revenue. Several were priced at $100M–$2B post-money anyway. The size isn’t the story. The timing is.

1. Born huge: the first real round is already enormous

In a healthy early market, a seed round buys “an idea and a team,” and valuation climbs round by round as milestones are hit. Here, that ladder is gone. One company took a ~$100M “seed++” that is essentially its entire funding to date — it has raised once, and once was nine figures. Another raised a $220M Series A at a $2B post with no prior scaled round.

Valuation has decoupled from milestones. The money doesn’t map to a shipped product; it maps to a founder’s position on a technical paradigm. Capital is making a bet: if this paradigm holds, this team is likely still at the table.

2. Round compression: the confirmation window collapses

The second signal is speed. When the anchor shifts from milestones to position, the cadence shifts from “advance on milestones” to “grab the slot before it closes.” Some companies have run the entire Angel-to-Series B sequence in months; round labels lag far behind the cash already in the door.

The counterintuitive part isn’t “they raised a lot”, it’s the collapse of the confirmation window. The interval between rounds is itself information: time for the company to hit milestones, and a window for the next investor to watch. Run that sequence in months and you’ve given the window up. Investors aren’t adding on more evidence; they’re deciding “wait any longer and I lose the slot.” Round compression is capital hedging missing information with speed.

3. Same paradigm bet, two playbooks

Here’s what only a cross-border view shows. The US and China are buying different positions, with different money.

The US track is software-first, betting on proprietary data flywheels. The checks come from tech billionaires and strategics — Bezos (twice this week), Eric Schmidt, Jeff Dean, Amazon, AMD Ventures, GV, In-Q-Tel. These companies will be priced as “data assets.”

The China track is hardware-first, betting on high-value real-world scenes — hotels and pharmacy retail, e-commerce logistics, 3C and auto manufacturing. The checks come from industrial and manufacturing capital — Wanxiang, BAIC Capital, Didi, Junpu. These will be priced on “scene penetration.”

Three years out, the same “physical AI” label lands in two different valuation models. Benchmark them with one ruler and you’ll be wrong.

Why now

Three forces converged. The LLM dividend window is closing, the strongest models are now free and open-weight, so capital is hunting the next “big and early” paradigm. A cohort of paradigm-grade founders was released from big labs and last-gen unicorns in 2025–2026. And two years of LLM investing proved that hesitating at a paradigm turn costs far more than overpaying one round.

What it means

For founders without a big-lab pedigree: the realistic path is to lock a high-value vertical scene and build data assets others can’t reach, not to fight for the general foundation layer. And remember that a high birth valuation is a liability, not an asset. It raises the bar for every round that follows.

For investors: the funds that judged by model benchmark scores now need a new ruler. Will this paradigm be falsified in three years, is this team truly at its core, and is the path to proprietary data or scenes already locked by an industrial player? In China especially, the best embodied targets may be claimed by industry capital before financial investors even see them.

For major tech companies and platforms: You are both the launchpad for these teams and potential buyers of them in the future. The outflow of top talents will continue, so you need to figure out in advance whether to build internal teams on your own or buy back these startups later.

For enterprise clients: Do not let huge financing rounds sway your purchasing decisions in the short term. Most of these startups are not yet capable of large-scale delivery, and their valuation does not equal their actual delivery capacity.

Tracking over 90 days

Median Round Interval: Tracks the time gap between the current and previous financing rounds for World Model / Physical AI track. If the median shortens to under 4 months, “round compression” is established as a structural trend; if it rebounds above 6 months, the heated financing is more likely an isolated phenomenon rather than a sustained trend.

Proportion of Unprofitable Companies with High Valuations: Calculates the share of Physical AI firms with a valuation of no less than $500 million and no scalable revenue in the current quarter. A rising proportion confirms “anchor shift”, while a falling proportion indicates the market is refocusing on operational milestones.

Sino-US Track Divergence Verification Rate: Monitors counterexamples such as “US enterprises shifting focus to scenario implementation” or “Chinese enterprises prioritizing data flywheel development”. Fewer counterexamples mean the judgment of “divergent development tracks between China and the US” is more credible.

Belief, not fundamentals. NewLimit, co-founded by Coinbase's Brian Armstrong, went from an $810M valuation to $3.1B in just 13 months — entirely before any clinical data. The money holding up that price is a specific kind: crypto and AI exit wealth. Armstrong put in ~$110M himself; the same cluster includes Sam Altman's Retro Biosciences ($180M seed, raising $1B) and Bezos-backed Altos Labs ($3B at launch). They are betting on "reversing aging" as a thesis, not on a de-risked asset.

The science and the moat. NewLimit does partial epigenetic reprogramming: using transcription factors to dial back a cell's aging markers while preserving its mature identity, deliberately avoiding the cancer-prone full Yamanaka route. The real moat isn't a single drug — it's a flywheel. Its Ambrosia AI platform trained on 6.5M T-cell records and finds working factor combinations at over 2x the baseline hit rate; its in-house pxbc-seq cuts screening cost by 100x; mRNA delivery runs about $1.20 a dose. Science is moving fast — its first liver-cell candidate arrived two years ahead of plan — but there is still only one preclinical candidate, and the entire clinic lies ahead.

The mirror: China's pragmatic answer. Per SVTR's AI venture database, life sciences is the most active AI application sector, yet the US dominates 341 deals to China's 15. In the cell-reprogramming cluster, China is essentially one company: Insilico Medicine, which runs AI drug discovery, has 28 pipeline assets (nearly half already in the clinic), and signed a deal worth up to $2.75B with Eli Lilly in March 2026. Tellingly, Lilly is hedging both models — NewLimit's preclinical reprogramming in the US, Insilico's clinical pipeline in China.

Why it matters for cross-border investors. Under the same "longevity" label, the pricing anchor differs entirely: the US bets on a scientific inflection point, with valuation running ahead of data; China ties valuation to clinical progress. A US framework will misprice the Chinese names, and vice versa. NewLimit is best read as a ten-year, high-uncertainty bet on belief — not a company you can value on fundamentals today.

Sources: SVTR AI venture database · Contrary Research · public funding announcements.

For three years, the assumed bottleneck in AI moved in a clean sequence: data, then model capability, then GPUs. SVTR’s latest weekly issue suggests it has moved one layer deeper — into electricity itself.

In SVTR’s AI Venture Database, Issue #161 (80 global AI rounds, June 2026), some of the largest checks went not to model or chip companies, but to companies answering a physical question: where does the power come from, and where does the heat go?

A caveat we hold ourselves to: SVTR’s weekly issue is a curated sample, not a market-wide census. But ~46% of this issue’s rounds (37 of 80) went to physical-world hardtech, and the direction is consistent enough to flag early. The smallest slice — energy, just three deals — is where the founder profile is changing first. That’s the signal.

1. Capital is funding power generation, not smarter models

Three companies that literally generate power raised large rounds in the same window:

SUNUP Fusion — compact controlled fusion on a deuterium–helium-3 fuel cycle; $100M Series A+.

Endurance Energy — deep-sea geothermal tapping heat near plate boundaries; $54M Series A led by Founders Fund.

Weilan Nuenergy Co., Ltd.，small modular reactors (SMRs, pressurized-water design) purpose-built for AI compute; hundreds of millions of RMB across angel+ and Pre-A.

What unites them isn’t a technology roadmap. It’s what they sell: baseload power. Not peak capacity, not a green label — 24/7, predictable, weather-independent electricity. That is precisely what intermittent renewables can’t deliver and what AI data centers most need. When an industry starts paying a premium for baseload rather than peak compute, its binding constraint has shifted a layer.

2. The founders changed: from model researchers to nuclear engineers

If you only look at the dollars, this reads as “energy is hot again.” The deeper signal is in who is starting these companies.

Weilan Nuenergy’s founder spent 20+ years in nuclear engineering, trained at Tsinghua.

SUNUP Fusion’s founder is a Fudan University physics professor with a plasma-physics PhD from UCSD and years running a fusion science institute.

Endurance Energy’s Andrew Redd was a life-support systems engineer at SpaceX, working on Dragon and Starship.

When a field’s top talent migrates from “models and algorithms” toward nuclear engineering, fusion physics and aerospace systems, the genuinely hard, genuinely scarce part of the problem has moved from bits to atoms. Model capability is commoditizing — open weights and APIs turn “smart” into a purchasable input. Getting a reactor onto the grid, or pulling stable geothermal energy from the deep sea, takes a decade of engineering and regulatory relationships. That is the moat software teams can’t replicate on a funding cycle.

3. Power and cooling now define siting — and the US and China diverge

If the scarce input is physical electricity, then where you can get power and where you can dump heat becomes a competitive variable, not a logistics footnote. Two companies in this issue treat siting and cooling as the product itself: Huahong builds fully liquid-cooled edge supercomputing, turning heat dissipation from a data-center constraint into a deployable product; Orbital puts GPUs on satellites, running inference in orbit where power and cooling conditions differ entirely.

The sharper split is in how the two largest markets solve supply. China leans nuclear — fission (Weilan Nuenergy’s SMRs) plus fusion (SUNUP Fusion) — backed by university engineering pipelines and patient state capital. The US is more dispersed — geothermal (Endurance), orbital (Orbital) — backed by dollar funds like Founders Fund and a16z, and talent spilling out of SpaceX and aerospace. For a cross-border investor, “investing in AI’s electricity” is two entirely different underwriting problems.

Why now

The shift is happening now, not earlier, because AI’s load profile changed. Training is a periodic peak load; inference is a continuous baseload. As the industry pivots from training a bigger model to running the model at scale, data-center demand turns from intermittent into a hard 24/7 requirement — and that collides with the grid’s physical limits. A data center takes about two years to build; a new transmission line or a new stable power source takes five to ten. Model gains are slowing and commoditizing while inference demand climbs; the two curves diverging pushed electricity from a background variable to the center of pricing.

Who’s affected

Investors: “powering AI” is becoming its own asset class, but its underwriting looks nothing like SaaS — multi-year timelines, heavy regulation, high capital intensity, exits closer to energy infrastructure than software. Pricing a power plant on software multiples is the easiest mistake for infra funds and software-AI VCs to make. Industrial and deep-tech capital that can read “deliverable, stable megawatts” will capture returns others can’t price.

Founders: at the application layer, inference electricity is becoming a real line in the cost structure, and over time it sets a floor under gross margin. If you’re building infrastructure, the moat is less about technical novelty and more about the ability to secure power — siting, grid connection, long-term power-purchase agreements, regulatory relationships.

Hyperscalers: they’re locking up baseload with long-term PPAs, which both creates demand for independent power and caps it.

Enterprise buyers: remember the long-run conclusion — the floor on inference pricing is set by electricity, not by models.

What we’re tracking (next 90 days)

1. Density and size of energy rounds. Whether new generation deals (nuclear / geothermal / fusion) explicitly tied to AI compute, at $50M+ per round, keep appearing, and whether the median check rises.

2. PPAs and siting. The count of long-term power-purchase or co-location agreements between hyperscale data centers and independent baseload sources (especially nuclear and geothermal).

3. Founder backgrounds. In newly formed “AI infrastructure” companies, the ratio of energy / nuclear / aerospace founders to model-background founders.

SVTR’s AI Venture Database now carries an AI-Energy tag, filter by generation path (fission SMR / fusion / geothermal), cooling and siting (liquid / orbital), and US–China comparison. If you’re evaluating a “power for AI” deal, this taxonomy locates its comparables and where the real risk sits.

On June 5, the AI-tonomy Summit 2026 was held at the headquarters of Plug and Play in Silicon Valley. More than 500 founders, investors and corporate executives attended the event. As a community partner of the summit, SVTR was present throughout the whole event. While most AI conferences over the past two years have focused on competing over AI models, this summit has seen a shift in its agenda focus.

This article will not recap the agenda. Instead, it extracts industrial trends and capital implications observed on-site, tailored for early-stage investors and founders who focus on the commercialization of agents and AI infrastructure.

There were more than 40 speeches throughout the day and 3 parallel sub-forums focusing on enterprise agent implementation, cutting-edge research, and modern AI infrastructure. No one was seriously discussing the countdown to AGI anymore; instead, everyone was talking about one thing: how to truly deploy AI into production systems. To sum up: the first phase of AI competition centers on model capabilities, while the second phase hinges on system implementation.

At the infrastructure roundtable, investors and entrepreneurs including Etienne Dilocker, CTO of Weaviate, Madison Faulkner from NEA, and Kyle McNulty from IQT reached a strong consensus: training remains important, yet the next major battleground lies in inference. Whether enterprises can adopt AI at scale ultimately depends on whether inference costs can be kept affordable, latency can be controlled for production use, and accuracy can meet commercial standards. The decline in token costs has fallen short of expectations, making the inference optimization stack composed of small models, fine-tuning, routing and caching far more crucial than anticipated. Many participants pointed out directly that less than 10% of enterprises have achieved large-scale AI deployment, with major bottlenecks existing in data, access permissions, compliance and governance.

When it comes to value redistribution: model companies captured nearly all the attention in the previous stage, whereas the next wave of dividends will likely go to four key sectors: inference, payment and identity, connectivity and distribution, as well as security.

The Four Layers of Infrastructure

1. Inference Layer: Turning “Adaptive Inference” into a Product

The inference layer has already produced deployable product samples. A representative case is Fastino Labs’ adaptive inference API product Pioneer, which in May 2025 closed a $17.5M seed round led by Khosla Ventures, bringing total funding to $25M. In April 2026, the company released an open-source fine-tuning and deployment package supporting Qwen, Gemma, Llama, and other major open-source models — enabling adaptive inference, automatic model iteration, and live dynamic upgrades. The model a user deploys on day one is the worst version they will ever use.

According to the SVTR database, inference is one of the few infrastructure sectors where independent companies have emerged separately in both the US and China:

• US: Cerebras, Groq, Fireworks, Baseten, Fastino

• China: MOFFETT AI, Enflame, SiliconFlow, QingCheng.ai, and more than ten others working on inference chips or inference stack optimization

This is a track widely favored by cross-border investors.

2. Payment & Identity Layer: Can an Agent Complete a Real Transaction?

The inference layer determines whether an agent can run. The payment and identity layer determines whether an agent can safely complete a transaction — being verified, accountable, and authorized. This is the core bottleneck in landing AI applications: many agents can already compare prices and generate order drafts, yet get stuck at the final payment step.

Two implementation directions have already produced working samples:

Application-layer sample — Snaplii:

• 350,000+ North American users

• 500+ partner brands

• Annual transaction volume exceeding $100M

• 1,700+ agents connected

• Its A2M Skill is natively MCP-compatible and can be directly invoked by large-model clients such as Claude

• It has implemented three principles of agent payment: one-time credentials, tokenization (the agent only sees the token, not the card number), and commercial-grade spend limits and rules that can be revoked at any time

Protocol-layer sample — Verisense:

• Focuses on discovery and interoperability between agents

• Each agent has a verifiable identity

• Behaviors between agents are verifiably recorded into a settlement system

Its development aligns with the latest industry trends. Google's AP2 protocol was released in September 2025 and donated to the FIDO Alliance in April 2026, with the newly added capability of unattended autonomous payment. Ping Identity's Identity for AI became officially available worldwide at the end of March 2026. Protocol-level projects feature long development cycles and rely on cross-ecosystem collaboration. They play a vital role yet have the longest verification periods among all project types.

US-China contrast: Companies in the payment and identity layer are almost entirely concentrated in the US, Europe, and Israel. Independent Chinese startups are largely absent. SVTR judges that domestic agent payments in China will most likely be embedded directly into Alipay and WeChat Pay. The opportunity for Chinese entrepreneurs lies in providing cross-border settlement interlayer services for agents going overseas.

3. Distribution & Security: One is still in the early stage, while the other is quite certain.

1) Distribution layer — still in its early stage

The cognitive sample here is Sense Space, whose thesis is that the “connection effect” of the internet is shifting toward an “awareness effect“ in the AI era:

• In the past, the internet distributed content to people

• In the AI era, capabilities are distributed to agents

This will give rise to new entry points for agent discovery, orchestration, and contextual recommendation. Based on its roadshow materials, this represents a platform revenue potential of over $400 million and an intermediary GMV exceeding $80 billion. However, the prerequisite for realizing this is that the agent population truly explodes.

2) Security layer — demand is already very clear

According to the founder of Horizon3.ai, attacks targeting AI systems have already emerged. Traditional security risks are amplified in the agent era, including:

• Being induced to invoke the wrong tools

• Permission abuse

• Information leakage during inference

We believe that in the coming years, AI-native security will evolve from an auxiliary module into a prerequisite for enterprises to implement AI solutions, and thus become a relatively definite budget priority for corporate procurement.

4. The Hidden Thread and Four Tracking Directions

Beyond the four infrastructure layers lies an industry hidden thread: as the cost of acquiring information through AI keeps falling, what is truly scarce is information credibility — credible people, credible circles, and credible project screening. This is also the core logic behind SVTR’s build-out of a cross-border AI venture capital platform.

In the next phase, SVTR will track four types of companies along these infrastructure layers:

• Inference layer: We focus on unit economics, as demonstration effects are not indicative of actual performance.

• Payment & identity layer: Distinguish two paths with completely different verification rhythms for applications and protocols;

• Connection & distribution layer: Track the real inflection point of explosive agent growth.

• Security layer: Prioritize coverage as a deterministic budget direction.

Those who can build AI with lower-cost inference, safer execution, smoother payment processing and more reliable collaboration hold greater advantages today. The edge of merely pitching AGI concepts is fading away.

From SVTR’s perspective, the most noteworthy takeaway from the latest round of financing is not how many Agent startups have secured funding, but the emergence of an entire cohort of firms dedicated exclusively to Agent governance, whose valuation timelines run well ahead of actual enterprise-level Agent deployments. While the market is still debating how many jobs Agents can replace, investors are already pouring capital into addressing a separate question: once Agents are fully operational, who will monitor their activity, restrict their access permissions, and shut them down in the event of malfunctions.

In short, while the practical capabilities of Agents remain under scrutiny, their inherent risk of going rogue has already been assigned a monetary value. This new intermediate management layer tasked with monitoring, security governance, workflow orchestration and activity logging is evolving into a standalone, revenue-generating infrastructure segment. Its core trait is that it leans more heavily into pure software architecture than execution-layer Agent products, enabling these businesses to command valuation multiples typical of pure software firms. Within the current financing landscape, most general-purpose productivity Agent projects are stuck at small seed-stage funding rounds, whereas several Agent governance startups have hit post-money valuations nearing $200 million at their Series A financing. We previously analyzed vertical Agent tracks in our last briefing—such as Sierra applying software valuation multiples to service-focused business models—and will not revisit that detail here. To draw a clear distinction: our prior coverage focused on which roles Agents displace, while this analysis centers on the third-party entities tasked with overseeing Agents.

Signal 1 | A full cohort of firms focusing on governance has emerged within a single funding batch, while corresponding execution-layer deployment has yet to gain meaningful traction.

First, examine the concentration of these enterprises.

• Gray Swan: Develops adversarial security testing and protection frameworks for models and Agents; secured a $40 million Series A round. Its core team hails from CMU, with founder Zico Kolter, a widely cited scholar in AI safety research.

• Geordie AI: Monitors Agent behaviors and access permissions within enterprise systems; closed a $30 million Series A financing, valuing the company at approximately $180 million post-money, with founding talent sourced from Darktrace.

• Canyon Code: Specializes in orchestrating interactions and dependency relationships between distinct Agents, having raised $5 million in funding.

• Modiqo: Converts validated Agent execution logic into reusable workflow modules, with $3 million in raised capital.

• Trajectory: Conducts ongoing model retraining using real-world end-user interaction data; secured a $15 million seed round resulting in a $115 million post-money valuation, with investment participation from Jeff Dean and Fei-Fei Li.

Adding Tekst, which builds process intelligence solutions, and Tensormesh, focused on cutting inference costs, nearly ten companies in this same funding cohort center their businesses on enabling observability, constraint management and reusability for Agents.

What stands out is this counterintuitive market dynamic: the enterprise-grade Agents these companies are built to support have not yet been widely deployed across industries. The market is building out full-suite management tooling for a technology still in limited mainstream adoption, reversing the typical industry development sequence. Traditionally, a technology scales into widespread use first, operational failures surface thereafter, and governance demands arise reactively in response. By contrast, funding is flowing nearly concurrently to governance-focused and execution-layer Agent developers in this cycle, with financing for certain governance verticals landing even ahead of mainstream Agent rollout.

Here comes the question: What exactly is being purchased with this money if the implementation team has not yet achieved scale growth?

Signal 2 | What enterprises pay upfront to purchase is “controllability” rather than “capability”

The answer lies within clients’ budget structures. The initial budget that enterprises are ready to spend on Agents is shifting focus from “how much work it can accomplish” to “whether it will behave erratically”.

The valuation of Geordie AI best illustrates this trend. Focused on Agent behavior monitoring, permission auditing and risk interception, the firm secured a post-money valuation of approximately 180 million US dollars at its Series A funding round, even though the Agents it monitors have not yet been fully deployed within most enterprises. Investors set such a valuation not based on the value generated by Agents, but on the potential costs stemming from unregulated Agent behaviors. A system capable of stopping Agents from corrupting data, conducting unauthorized operations or leaking confidential information before such incidents occur matters to enterprises not for “efficiency improvement”, but for enabling “confidence in adoption”. Security and observability have evolved into the critical toggle that moves an Agent from PoC to formal production deployment, instead of supplementary fixes implemented post-launch.

Gray Swan presents the same market shift from another perspective. It specializes in pre-launch adversarial testing and red team exercises, turning the verification of “whether an Agent can be manipulated into malicious conduct” into an outsourced service. This business model carries a clear implication: enterprises acknowledge their inability to fully govern Agents in-house and therefore outsource the associated governance risks. This clearly proves that controllability has transitioned from an internal responsibility of engineering teams to an independent procurement category.

Once controllability becomes a standalone paid service, a more fundamental question emerges: given its separability in pricing, does it still follow the same valuation framework as core execution capabilities?

Signal 3｜A divide is emerging between the governance layer and the execution layer, with the former poised to command software-style valuations

This is the most risky yet high-conviction call in this analysis. Valuation rationales for the governance layer and execution layer are diverging.

As pointed out in the previous analysis, the execution layer (vertical Agents in particular) faces an awkward economic model: it replaces professional services, featuring the gross margin structure and delivery costs typical of service-based businesses, yet frequently gets valued as software. The governance layer operates differently. Functions including monitoring, security, orchestration and retraining are inherently horizontal, standardizable and reusable as the number of Agents expands. The marginal cost of onboarding the 100th Agent onto one monitoring system is drastically lower than that of the first, a defining characteristic of genuine software.

This will drive a valuation reversal: market participants assume software valuation multiples belong to task-performing Agents, but in reality, enterprises that oversee Agent operations may be the true beneficiaries of such multiples. Trajectory secured a post-money valuation of $115 million at its seed funding round not by executing specific tasks, but by occupying the horizontal niche of continuous model improvement; every real-world user interaction from new clients fuels and reinforces its unified feedback flywheel. Tensormesh follows a comparable logic: it does not build Agents itself but drives down operational costs for all existing Agents.

This dividing line merits close ongoing monitoring. If this trend holds true, capital markets will increasingly assign software valuation premiums to the horizontal Agent management layer over the near term, while pricing task-focused vertical Agents against metrics aligned with traditional service industries. Put simply, the most software-like segment within the Agent economy may not be the Agent products themselves.

Causes | Why Now

Why is this signal emerging at this moment instead of earlier or later? Three developments have converged within the same timeframe.

First, Agents have evolved from “being able to chat” to “being able to take practical action”, marking a qualitative shift in capability boundaries. Capable of tool invocation, data modification and autonomous sequential decision-making, such Agents now carry concrete and costly risks if out of control, turning governance from an optional add-on into an indispensable prerequisite for deployment.

Second, actual cases of unauthorized operation and malfunctions from leading Agent developers have circulated among enterprises, shifting buyers’ priority from evaluating functional merits to first enquiring about contingency plans for potential failures.

Third, having suffered valuation losses on vertical Agent businesses due to inappropriate software-style pricing for service-oriented offerings, capital investors are proactively hunting for software-like segments across the Agent industrial chain, with governance-related sectors successfully attracting this redirected investment.

Essentially, the current market momentum hinges on investment betting not on Agents’ present capabilities, but their future potential.

Impact | Who Will Be Affected

For entrepreneurs, a clear market positioning window has emerged. If you are building execution-layer Agents, one critical pitfall to watch out for: your core value may not lie in the actual tasks you complete for clients, but in your ability to deliver observable and controllable mechanisms that convince clients to deploy your Agent into production environments. Embedding controllability as an intrinsic product feature rather than an afterthought for compliance directly determines the valuation multiple you can secure in your next financing round. For those developing governance-layer products, sustainable competitive moats do not stem from isolated functional modules—monitoring, testing and orchestration features can all be quickly replicated. Instead, the true edge lies in evolving into a horizontal standard layer that becomes increasingly irreplaceable as more Agents connect to it and more data accumulates; this is exactly the market positioning contested by Trajectory and Geordie.

For investors, this calls for two distinct valuation methodologies. Higher software valuation multiples are justifiable for governance-layer and feedback-layer enterprises thanks to their robust reusability and tangible network effects. In contrast, vertical execution-layer Agent businesses need to be valued strictly based on delivery costs and gross profit margins, and investors should avoid overpaying using software multiples lured by the narrative of AI replacing professional services. A practical due diligence step involves splitting an Agent firm’s revenue into two buckets: income generated from task execution and income from risk assurance services. The proportion of revenue from the latter often serves as a better indicator of the company’s long-term fair valuation than its revenue growth rate.

For large platform operators, foundational model providers will likely integrate core monitoring, logging and access permission functionalities natively into their platforms. Independent governance vendors therefore must pursue deeper technical depth, broader industry coverage and stricter neutrality to sustain viable market space. For enterprise buyers, the primary evaluation benchmark for Agent procurement is shifting from capability demos to security and observability requirements, which will be included in procurement checklists as early as the PoC phase.

Verification: Three Core Metrics Worth Tracking

Over the next 90 days, validate this judgment via three quantifiable metrics.

First, the valuation multiple gap between governance-layer and execution-layer businesses. Track the financing multiples of companies focused on the new round of governance-layer functions (monitoring, security, orchestration and retraining), and calculate their multiple differential against general or vertical execution-layer Agents over the same period. A sustained widening of the gap confirms industry divergence.

Second, the priority ranking of “security and observability” on corporate Agent procurement lists. Based on publicly available enterprise AI procurement cases and RFPs, tally the percentage of instances where “observability, security and compliance” is specified as a mandatory prerequisite instead of an optional add-on. A notable uptick within 90 days validates the second indicator.

Third, the disparity in the number of financing deals for Agent governance-layer ventures between China and the US. Compare newly closed financing rounds targeting Agent governance-layer (excluding execution-layer) businesses in the US, UK and China across the 90-day window. If US-backed deal volume retains its lead with no narrowing gap, the Sino-US structural divergence outlined in the first indicator is substantiated.

Continuing to evaluate agent companies with standard SaaS metrics — LTV/CAC, net retention rate, 120% expansion rate — will systematically underestimate their growth ceiling: global enterprise IT budgets sit at roughly $500 billion; human-related spending (wages, professional services, outsourcing, recruitment costs) is more than ten times that. The underlying AI infrastructure layer is a separate discussion beyond this issue’s scope.

Signal 1 | Professional Domain ARR Growth Already Differs by Order of Magnitude from SaaS

Several data points observable over the past month:

• Contrario, a YC W25 AI recruiting company, reached $6M ARR in under 6 months on a seed round of just $2.3M

• Viktor, an AI coworker embedded in Slack and Microsoft Teams, hit $15M annualized revenue in 10 weeks, raising a $75M Series A

• Searchable, an AI search visibility platform for brands, reached $2M ARR in a few months

• Vétir, an AI outfit selection and wardrobe management app, disclosed 3,500% B2B customer growth and 9x annual revenue

Benchmark: Traditional SaaS companies take a median of 3–5 years to go from zero to $10M ARR. These AI agent companies did it in months.

The difference isn’t product quality — it’s who’s making the buying decision. Contrario’s buyer isn’t the CIO; it’s the HR director. Viktor’s buyer isn’t IT ops; it’s the business team lead. When an HR director is calculating “can this tool save me the cost of a $120K/year recruiting manager,” the decision logic is no longer software procurement — it’s labor substitution. The former has a small budget pool and slow decisions; the latter has a large budget pool and fast decisions.

This leads to the more important question: What exactly are investors pricing in right now?

Signal 2 | Capital Is Systematically Valuing “Professional Services Substitutes”

Reorganizing mid-to-late-stage financings from the past month by the function being replaced:

• Legal: Stilta ($10.5M seed, a16z + YC, AI patent analysis), Lexroom ($50M Series B, European civil law AI law firm operations), LawX ($8.7M seed, AI legal case sourcing system)

• HR / Recruiting: Contrario ($2.3M seed, $6M ARR in 6 months)

• Wealth & Asset Management: Farther ($150M Series D, unicorn status, AUM exceeds $ 23B), Moment ($78M Series C, AUM exceeds $ 10tn)

• Medical Administration & RCM: Commure ($70M raise, $7B valuation)

• Email Security & Enterprise Communications: Ocean ($28M raise, processing over 1 billion emails monthly), Tribal ($10M seed, agents on Salesforce)

These companies boast high valuations when measured by SaaS metrics. However, if we adopt a different benchmark — the total global payroll for the professional roles being replaced — their valuations actually appear conservative. Farther targets the service fee income of global wealth advisors; Lexroom cuts down the billable hours spent on legal research by European law firms; Commure reduces the labor costs of entire teams responsible for revenue cycle management in hospitals.

One detail worth noting: Farther has completed its Series D round, Lexroom its Series B, and Stilta its seed round. Yet their valuation-to-ARR ratios have converged. This indicates that investors believe this track boasts a high success rate and clear exit prospects, so they are willing to offer late-stage valuations for early-stage companies. This represents a capital trajectory entirely different from that of SaaS businesses.

So are the moats of traditional SaaS (product complexity, integration depth, ecosystem dependency) still important? The answer lies in the profiles of the founders.

Signal 3 | Founder Profiles Have Shifted from “SaaS Serial Entrepreneurs” to “Core AI Engineering + Domain Expertise in Parallel”

Representative founder backgrounds from recent financings:

• Hark: Brett Adcock, current CEO of Figure; serial entrepreneur (Vettery acquired by Adecco, Archer went public). New company raised $700M Series A at $6B valuation

• Decart: Dean Leytensdorf, Israeli 8200 military intelligence unit; earned a Technion CS PhD at 23. $300M raised, ~$4B valuation

• Unframe: Shay Levi, co-founder of Noname Security (acquired by Akamai for $450M), second-time founder

• IrisGo: Jeffrey Lai, former Apple Siri engineering manager, incubated by Andrew Ng’s AI Fund

• Modal Labs: Erik Bernhansson, former CTO of Better.com, former engineering manager at Spotify

• Catena Labs: Sean Neville, co-founder of Circle

These founders share one trait: they aren’t people who “know SaaS GTM” — they’re people who “understand core AI capabilities + have judgment in a specific professional domain.” Investors can complete due diligence faster; the resume accounts for half of the process. Early-stage valuations in the agent space are driven up by founder scarcity, not product/revenue metrics. This in itself is a fork in the road from typical SaaS investment logic.

Cause | Why Now

Three variables converged simultaneously from late 2024 to mid-2026, creating the current market:

1. Technology matured. LLMs crossed the reliability threshold for independently completing professional tasks. Repeatable work — legal research, recruiting screening, wealth advisor queries — now has stable product forms.

2. Budget shifted. Beyond enterprise IT budgets, the “human labor budgets” of HR, operations, legal, and finance departments are now formally part of AI procurement discussions. The decision-makers for this budget pool are not in IT — they’re on the front lines of business.

3. Distribution matured. Salesforce, Slack, and Microsoft 365 have opened up agent integration channels. Agent products no longer need to build their own sales and distribution paths — they can embed into existing enterprise entry points. Distribution factors now outweigh product in weighting. For example, Viktor’s ability to reach $15M ARR in 10 weeks owes far more to distribution than to product.

These three things had each happened individually before. Their simultaneous occurrence is what constitutes a brand-new market structure.

Impact | Who Gets Affected

Entrepreneurs: The window for vertical professional domains is shorter than commonly assumed. Don’t build generalized “agent platforms” — beware of being directly displaced. If entering a horizontal track, don’t build a general platform; be the #1 in a vertical domain, owning one function or one job type. The key shift is pricing from “per seat” to “per labor hour / task saved.” Building a general-purpose agent platform has high failure probability, because agents are more scenario-specific and fragmented than SaaS, and the channel economics above them are already forming. One contrarian view: model capability is not the biggest moat — “being able to clearly calculate how many labor hours I save clients each month” is what determines whether you win. Customers will pay a premium for quantifiable human cost savings.

Investors: The valuation anchor needs to change. LTV/CAC still matters, but the ceiling should shift from “this SaaS sub-sector market share” to “total global wage expenditures of the corresponding function.” The core of net retention should shift from “seat expansion rate” to “labor substitution growth rate“ — customer stickiness is no longer a pleasant surprise; what matters is human market share for that function. Agent company return cycles will compress from SaaS’s 7–10 years to 5–7 years, as valuation jumps happen faster and both PE takeover and strategic acquisition exit paths open simultaneously. Early ARR should also be re-framed around “payroll substitution”: customers care more about how much you save them. As long as agent task reliability exceeds what’s promised, retention cycles will be shorter than SaaS but customer expansion will be higher.

Big Platforms: Salesforce, Microsoft, Slack — companies with workflow entry points — are acquiring new strategic leverage. The rules they set for who gets agent channel access will themselves become the source of the next generation of take rates.

Enterprise Buyers: The budget boundary between CIO and CHRO/CFO is beginning to blur. A hybrid procurement committee will emerge in the next 12 months.

Validation: Three Key Metrics to Track

• Within 90 days: Track whether representative companies like Contrario, Viktor, and Stilta continue maintaining ARR growth exceeding 5x annualized rate. Regression to the SaaS median (2–3x) would mean the “payroll substitution” narrative has been over-discounted by the market.

• Within 90 days: Watch for whether any agent company formally writes a “percentage share of saved labor budget” clause into contracts. This is hard evidence that the labor-hour billing model is formally accepted — and will trigger enterprise finance-side accounting classification changes.

• Within 90 days: Watch whether Salesforce or Microsoft launches a clear take-rate mechanism for agent integration (like an App Store model). Once established, channel economics will rapidly reshape early-stage valuations. Horizontal agent companies bound to major platforms will see their valuations surge first.

The timing of the filing was not coincidental: on May 18, a California federal jury dismissed all of Elon Musk’s claims against OpenAI, clearing the most critical legal obstacle in the IPO process. In September, CFO Sarah Friar favored a longer preparation timeline. The current filing reflects CEO Sam Altman’s push for a September listing, with the CEO’s position prevailing over internal disagreements on pacing.

The article argues that the real signal of this filing is not the label of “world’s largest AI IPO”, but rather that OpenAI chose to proactively lock in public market pricing power and narrative control at the moment when the three battlegrounds — ARR growth rate, operational profitability, and valuation multiples — are about to fully diverge from Anthropic. This is a defensive preemption, not a winner’s posture.

I. ARR Growth and Valuation Multiples Have Already Reversed

When the annualized revenue trajectories of both leading companies are plotted together, the reversal inflection point appears in Q1 2026:

• OpenAI rose from $20 billion ARR at end-2025 to approximately $24 billion by April — ~25% growth in 5 months;

• Anthropic surged from $9 billion to $30 billion (official disclosure) over the same period, and according to follow-up reports from The Information and SemiAnalysis, the run rate was approaching $44 billion by May.

• OpenAI Q1 actual revenue was $5.7 billion; Anthropic Q1 was $4.7 billion — the gap narrowing to a $1 billion range.

Valuation multiples calculated on current disclosed bases:

• OpenAI: $852B / $24B ARR ≈ 35x

• Anthropic: $380B / $14B ARR ≈ 27x in February funding round negotiations; $900B / $44B ARR ≈ 20.5x in this month’s new negotiations

The implied secondary market pricing for Anthropic has also touched $1 trillion. As Anthropic overtakes on ARR, valuation multiples are converging — meaning OpenAI can no longer sustain a 35–40x ARR multiple on the basis of “industry leader” scarcity premium. This is the direct trigger for the preemptive filing.

II. Is the $1 Trillion Anchor Sustainable?

Bull case for $1 trillion target valuation has three pillars:

1. Private valuation trajectory rising consistently for 12 months: ~$157B at end-2024 → ~$500B in October 2025 → $852B this round;

2. Microsoft remains both largest customer and largest compute provider — the partnership framework provides a capital markets narrative of locked-in revenue streams;

3. Dual consumer and enterprise lines — business structure is broader than API-only Anthropic.

Bear case data is equally clear:

1. Absolute growth has been overtaken: Anthropic grew ARR 4.9x in 5 months; OpenAI only 1.25x over the same period. Even at a 40x ARR target, OpenAI’s pricing is less justified than Anthropic’s current implied 20–22x multiple;

2. Operational leverage is moving in opposite directions: Anthropic may achieve its first operational profit in Q2; OpenAI’s losses are widening over the same period. CFO Friar — bound by a 5-year, $600 billion compute commitment — has publicly acknowledged that revenue growth is insufficient to support all data center construction. This is an unavoidable risk exposure in the S-1 roadshow;

3. Governance terms unresolved + banker conflicts: Revenue sharing with Microsoft, IP ownership, and board power dynamics are not fully publicly disclosed. The underwriting syndicate of Goldman Sachs and Morgan Stanley are simultaneously in talks with Anthropic about its IPO. The same roadshow team must pitch two direct competitors to the same pool of LPs — this is yet another motivation for OpenAI to file first.

III. Transmission to the financing environment of the track

OpenAI’s filing will squeeze competitors on two levels:

1. Infrastructure comparable anchor reset

GPU cloud and training platforms like CoreWeave, Lambda Labs, and Together AI — once detailed OpenAI compute expenditure data becomes available as a benchmark — will face recalibration of their current 9–15x revenue multiples in private market pricing. Some companies may be forced to prematurely disclose take-or-pay contract details.

2. Window compression effect

SpaceX publicly filed its S-1 on May 20 (target valuation $1.75 trillion, raising $75 billion, Nasdaq ticker SPCX), with a roadshow scheduled to launch June 5 and pricing around June 12. OpenAI confirmed its filing just two days after SpaceX’s S-1 went public — the time-window pressure is immediate. Q3–Q4 will see at least $150 billion combined in mega-issuances from SpaceX and OpenAI. The listing window for mid-tier AI companies is effectively pushed to Q1 2027 or later.

For Anthropic, however, OpenAI’s filing timing actually presents a tactical opportunity: its proposed $900 billion private round can be completed before OpenAI’s public S-1, positioning Anthropic as the “unlisted but ARR- and profit-stronger” comparable before its own roadshow.

IV. Three Variables to Watch Over the Next 1–2 Quarters

1. OpenAI’s public S-1 disclosure — gross margin and Microsoft revenue-sharing mechanism

The public S-1 is expected 15 days after the confidential filing, i.e., no earlier than mid-to-late August. Key metrics to watch: if gross margin is significantly below secondary market implied assumptions, or if the Microsoft revenue-sharing mechanism is above market-estimated take rates (~20% range), the $1 trillion valuation will likely fall back to the $750B–$850B range.

2. Whether Anthropic files before OpenAI’s public S-1, and its disclosed Q2 operating profit data

If Anthropic files and can demonstrate Q2 profitability, OpenAI will face the “industry leader” narrative being dismantled. OpenAI’s valuation would face substantive downside, with the lower bound potentially breaking through $700 billion.

3. SpaceX’s June offering results and final pricing range

SpaceX’s current target is $1.75 trillion at $75 billion raise. If the final issuance valuation falls below $1.5 trillion, or the price drops to the lower end of the issuance range within 30 days after pricing, it indicates that the market’s capacity to absorb trillion-dollar landmark IPOs is nearly saturated. This will directly affect the difficulty of securing anchor investors and the pricing ceiling for OpenAI’s roadshow in September.

On May 18th, SVTR partnered with Zhiwai to host a closed-door luncheon on “Chinese AI Brands Going Global” in Silicon Valley. The event gathered entrepreneurs building consumer brands, AI applications, cloud services, influencer marketing, e-commerce platforms, and deep-tech startups on US marketplaces.

Participant niches included: consumer brands for specific US demographics, AI-powered influencer marketing, UGC licensing, content platforms for micro-merchants, deep-tech ventures, and cloud infrastructure.

The core question everyone grappled with: As AI, supply chains, content channels, and global markets shift simultaneously, how should Chinese entrepreneurs rethink “brand globalization”?

Deep Needs Behind a Small Niche

After the event, Zhiwai founder Lina shared a representative case: An entrepreneur building an electric vehicle for the American hunting community.

At first glance this seems extremely niche. But digging deeper, the needs run surprisingly deep:

The US has a large population of licensed hunters who are highly willing to invest in gear. But standard e-bikes completely fail their real scenarios: climbing hills with gear, navigating muddy terrain, cold-weather starting, long-range needs, chain-break resistance, and high climbing ability.

What hunters actually need: a vehicle that handles 30-degree slopes, starts at -20°F, carries 200 lbs of gear, and lasts a full day on a single charge.

Even more noteworthy: the entrepreneur himself is the target user. The project started from “I’m fed up with this problem myself” — not “I’m going to build something for a big market.”

This case sparked the core discussion: Must great entrepreneurial opportunities start from “big markets”? The answer is no.

Often, a small circle actually has sharper pain points, more genuine demand, and stronger word-of-mouth — making it easier to carve out a true niche market. Small community, high pain point, strong reputation — this may well be the starting point for a global brand.

Especially in the US market, segmented user groups have very specific needs and clear willingness to pay: hunters, outdoor enthusiasts, micro-merchants, content creators, or professional users — they don’t need “everything to everyone.” They just need a tool that truly understands their scenario and solves their specific problems.

This is what entrepreneurs need to rethink: Going global isn’t just exporting domestic goods; it’s about using global supply chains, AI tools, and local user insights to redefine the product entirely.

AI’s Penetration into Brand Globalization

The session also covered AI’s real-world applications in traditional industries: Even in manufacturing like electric bicycles, AI has entered the product development process. The hunter e-bike startup is already using LLMs to assist with code writing, drivetrain design, and generating product mockups.

AI’s impact extends far beyond software and marketing — it has seeped into manufacturing and hardware entrepreneurship.

In the past, building a hardware brand required a long cycle: product definition, engineering design, supply chain coordination, market testing, and content distribution. Now AI helps teams move faster on solution design, prototype expression, content production, user feedback analysis, and market testing — dramatically improving efficiency.

But AI doesn’t replace an entrepreneur’s understanding of users. It amplifies the advantages of teams that truly understand users, scenarios, and supply chains.

Four Typical Paths for Chinese AI Brand Globalization

1. Leveraging China’s supply chain to serve niche US markets
Like the hunter e-bike — the essence isn’t “e-bike exports,” but redefining products around high-pain-point user groups.

2. Next-generation brand infrastructure powered by AI
Influencer marketing, creator discovery, video content analysis, and UGC licensing. Projects like Influcio, Chimeboard, and Vyrill help businesses better understand content, influencers, users, and conversion logic.

3. New store-opening platforms for micro-merchants
Aladdinmall focuses on online commerce, local discovery, pre-purchase communication, managed payments, and mobile shopping — lowering barriers for local businesses to go online.

4. Deep-tech ventures from foundational technology
Breaking through from the technology side to find industrial application scenarios for global markets.

AWS and startup solutions architects also contributed perspectives on cloud infrastructure, MVP architecture, AI agents, LLM application development, and modern data stacks — supporting startups as they scale from early product to growth.

All of this points to one trend: Brand globalization is no longer “selling Chinese goods overseas.” It has become a system-wide endeavor: product definition, supply chain organization, user insights, content marketing, channel building, AI tools, data infrastructure, and local trust networks.

For Chinese entrepreneurs, the US market is both a proving ground and an amplifier: real user needs, diverse niche markets, mature commercial infrastructure, and customers willing to pay for professional products. But it also demands that entrepreneurs truly understand local users — rather than equating “globalization” with cross-border sales.

Great brands often don’t start from a grand slogan, but from a real user’s real pain point.
A small niche may hide a deep need.
A specific user group may lead to a global brand.

This is the direction SVTR aims to connect: gathering practitioners in products, brands, technology, supply chains, traffic, and industrial services — letting real experience flow and helping more Chinese entrepreneurs discover real opportunities in global markets.

SVTR believes this is one of the most underrated secondary signals of 2026. A group of elites with the most solid information advantages are voting with their feet to redefine the real moat of the vertical Agent track. It lies neither in models nor data, but in category insights and go-to-market strategies. The homogenization at the model level has driven the marginal value of merely understanding technology down to nearly zero. Meanwhile, the genuine barriers of vertical Agents happen to fall within the core strengths of these VC partners.

It is noteworthy that such career shifts have not taken place in seemingly grander fields such as model development, foundational AI enterprises or chip sectors. Even the most high-profile tracks have failed to attract partners to step into actual business operations. As the window for information arbitrage on deal sourcing narrows, they opt to participate in the industry in person rather than making more aggressive investments.

Signal 1 | Partner migration is highly concentrated in the "enterprise back-end"

Over the past 12 months, there have been at least seven clearly documented cases where partners from mainstream US dollar funds left their positions to start their own businesses:

• a16z GP Anjney Midha → Amp: AI compute tuning, first-round raise exceeding 130M

• Kleiner Perkins GP Joublin Mirzadegan → Roadrunner: AI-native CPQ, with a total funding of 27 million US dollars raised through seed and Series A funding

• a16z Partner Jack Soslow → Ciridae: an AI operation system for mid-sized enterprises, which secured $20 million in seed funding

• Founders Fund Partner Sam Blond → Monaco: an AI sales platform, which raised $50 million in Series B funding with a valuation exceeding $500 million

• Basis Set Venture Partner Ian Crosby → Synthetic: an AI accounting firm, which obtained $10 million in seed funding

• Rough Draft VP Sahir Jaggi → Forus: an AI prescription and insurance process platform, which closed a $160 million initial round and became a unicorn enterprise

• Twitchex CSO, angel investor with investments in over 50 companies ColinCarrier → OriginLab: a provider of world model training data, which raised $8 million in seed funding.

If we broaden the scope to include senior investors and serial entrepreneurs, and further add Cowboy Space co-founded by Baiju Bhatt, co-founder of Robinhood (raised $275 million in Series B funding with a valuation of $2 billion), Exponent founded by Charlie Cheever, co-founder of Quora, and Recursive Superintelligence led by Richard Socher, former Chief Scientist of Salesforce (post-money valuation of $4.65 billion), the figures will rise even higher.

What makes this list counterintuitive is not the fact that these people have left their previous positions, but that they have collectively steered clear of the most high-profile tracks such as large models, chips and consumer-oriented AI.

Signal 2 | Not Betting on Technology, But on “GTM × Category Insight”

Analyzing what these companies are offering: CPQ solutions, accounting bookkeeping, sales compensation management, prescription pricing, franchisee financial management, insurance brokerage services, sales process management and more. Almost all of them are enterprise back-end businesses that are mundane yet deeply accumulated, entail high switching costs and require mastery of industry-specific terminology. The real barriers to entry in these sectors have never lain in model selection, but in the tripartite combination of industry expertise, client relationships and sales rhythm control.

What core strengths do VC partners possess? They have built up insights into customer profiling, sales rhythm mastery and valuation benchmarking by evaluating 2,000 companies over a decade. These are precisely the most scarce resources for vertical Agent companies in their rapid expansion phase. The biggest bottleneck for founders of vertical Agent enterprises is usually not how to train Agents, but how to get involved in procurement negotiations, set reasonable pricing benchmarks and secure the first batch of reference clients. Former general partners hold inherent structural advantages in these three aspects.

In other words, they did not enter the industry merely drawn by the AI boom; instead, they realized their inherent strengths could for the first time be monetized on the seller side. If such core strengths can be revalued, what changes will take place across the entire VC value chain?

Signal 3 | The Sourcing Era Is Ending; The Underwriting Era Is Just Beginning

This is truly a third-tier pivotal signal. If the optimal strategy of frontline GPs has shifted from “finding the next Forus” to “establishing the next Forus independently”, the value chain of the VC industry is being rewritten.

Over the past five years, excess returns from AI investments mainly stemmed from sourcing advantages — being the first to spot OpenAI, Anthropic, or issue term sheets to Cursor. However, vertical Agent projects are overly saturated and fragmented, and the technical gaps are too narrow for GPs to maintain prominent sourcing edges. The likelihood of five high-caliber AI Sales Agent startup teams emerging simultaneously is far higher than that of five equally strong foundational AI company teams.

As information arbitrage opportunities in sourcing fade away, excess returns will flow back to two types of participants: funds equipped with industrial know-how to conduct professional underwriting, and partners with GTM leverage capable of launching startups on their own. The logical closed loop behind this trend is fully formed. The only remaining question is: why did this trend break out intensively between 2025 and 2026, rather than in 2023 or 2024?

Causes | Why Now?

Three major forces are taking effect simultaneously:

1. Technological inflection point. The marginal capability gaps at the model level are narrowing to zero, forcing vertical Agents to shift their competitive moat toward go-to-market strategies and category insights. In 2024, startups could still secure Series A funding by claiming “superior Agent orchestration technology”, yet this growth narrative has lost its appeal by 2026.

2. LP inflection point. Starting from Q3 2025, dollar limited partners have grown far less patient with reinvesting in foundational AI companies, and capital is flowing toward application-layer ventures with proven revenue streams and verifiable unit economics. Leading general partners can directly sense this shift in sentiment during LPAC meetings, at least six months earlier than such trends appear in secondary research reports.

3. Incentive inflection point. The carry profit realization timeline of top-tier GPs is no longer aligned with the market opportunity windows they identify. If a GP forecasts a seed-stage project they would traditionally back could surge tenfold in value within five years, while carry profits locked in the fund take eight years to vest and get split among a 20-person core team, the expected returns from personal direct investments outpace long-term career gains within institutional funds.

This checklist comes into being precisely as these three inflection points converge.

Impact | Who Gets Hit

Founders: The profile of real competitors has shifted. Over the past 18 months, vertical Agent entrepreneurs have mainly faced competition from peer founders and incumbent software companies. Going forward, this competitive landscape will welcome a new group of “ex-GP founders” who come equipped with limited partners (LPs), go-to-market (GTM) networks, client rosters and secondary valuation benchmarks. Their edge lies not in products, but in pre-established financing rhythms and sales networks. They can mobilize over 20 million US dollars at the seed round stage, and skip traditional demos to directly connect with strategic LPs in Series A financing. Ordinary founders need to rethink their differentiated positioning: superior technology is no longer sufficient. They must prove they possess irreplicable industrial expertise in specific niche sectors that even GPs cannot match, such as exclusive client connections, unique data flywheels, or hard-to-replicate founder-market fit within the industry. A concrete strategic takeaway is that instead of pursuing general horizontal Agents, it is wiser to firmly anchor the business in vertical industries overlooked by GPs.

Investors: The era of deal-sourcing arbitrage is drawing to an end. A clear inference is that if even firm partners are launching their own startups, the scarcity of merely discovering investment deals will inevitably decline, while the scarcity of accurately evaluating projects will rise sharply. The moat of early-stage funds is shifting from gaining early access to projects to identifying industry ceiling potentials ahead of others. This means capabilities once regarded as cost centers, including investment research teams, industry analysts and exit return models, are becoming core factors in investment valuation. Meanwhile, the transferable value of secondary valuation benchmarks and LP connections is growing increasingly prominent, and early-stage personal connections alone can no longer sustain the differentiated competitiveness of new funds. Mid-sized funds need to stay highly vigilant; under this new supply-side logic, funds that are neither large enough to provide underwriting support nor close enough to participate directly in industrial layout are in the most precarious position.

Platforms and Major Enterprises: These vertical Agent enterprises pose the most direct substitution threat to SaaS giants including Salesforce, ServiceNow and Workday within the next 24 months, and merger and acquisition activities are expected to surge notably starting from the second half of 2026.

Enterprise Clients: CIOs and CFOs are no longer faced with the simple choice of whether to adopt AI tools. Instead, they have to decide between purchasing AI modules developed by established incumbents or brand-new AI-native platforms built from scratch. The latter have already accumulated client cases with annual recurring revenue (ARR) ranging from 600,000 to 1 million US dollars.

Verification: Three Key Indicators Worth Tracking

1. Within 90 days, check if more than two current partners from a16z, Founders Fund, Kleiner Perkins and Sequoia resign to start their own businesses, with their business focus still centered on enterprise backend Agents. If this happens, it indicates a structural industry shift rather than isolated individual cases.

2. Verify whether the median post-money valuation of seed-stage vertical Agent enterprises exceeds 50 million US dollars before Q3 2026. This is a core threshold to judge whether general partners are systematically lifting the bottom valuation of this track.

3. Confirm if there emerges the first public case of seed financing initiated by former GPs with participation from their former venture capital funds within 90 days. This will officially signal the recognition of this new supply-side logic by limited partner capital.