$5.2B+
Raised by pre-product labs
01 · The frontier board
Five problems, thirty-eight bets
Most AI companies pick a market. A small number pick a problem that no one has solved and then raise against the possibility that they will. This board tracks the second kind — sorted by the problem, not the company, because the problem is what tells you whether the bet is real.
7
Launched since April 2026
4
Turing / Nobel-tier founders
0
With a shipped product
How to tell a frontier bet from a wrapper
Every company on this board passes at least four of these five tests. It is a useful filter in either direction: apply it to a pitch deck and the ones that fail tend to be selling distribution, not research.
TEST 01Names the failure, not the marketThe pitch starts with something LLMs provably cannot do — not with a TAM slide. Sutton's Oak Lab opens by calling current AI fundamentally broken.
TEST 02Has run the experiment beforeFounders come from a lab bench, not a GTM org. The four Google researchers behind Discovery Loop shipped MapReduce, AlphaStar and seq2seq.
TEST 03Capital precedes productA $1.1B seed with no product is not a red flag here — it is the compute bill for a training run that has to happen before anyone can judge the idea.
TEST 04Has a falsifiable milestoneA dated benchmark or a public experiment. Ineffable promised first benchmarks by late 2026. Skyfall said it will buy a company and run it.
TEST 05Owns its data sourceGeneral Intuition trains on action-labelled gameplay from a platform it already owns. Renting the data is a strategy; owning the faucet is a moat.
Filter
Problem 01 · Self-improving systems
Can an AI system do AI research?
Every frontier lab now runs some internal version of this loop — models that help design the next model. Nobody has closed it. The bet on this row is that whoever automates the researcher, rather than the coder, compounds fastest and hardest.
Why it is still open: current systems can write code and run evals, but cannot choose a research direction, notice that a result is interesting, or decide what to try next. That judgement is the entire job.
| Company | Capital | The specific bet |
|---|---|---|
| Mirendil Aug newsEx-Anthropic · ~20 people | $200M seed a16z + Kleiner, Nvidia |
Sells the autoresearch platform the big labs build for themselves and guard. Founder Behnam Neyshabur worked on AI-for-science at Google and Anthropic; his co-founder built Anthropic's first internal autoresearch system. Signed a $100M+ multi-year Google Cloud deal on 6 August — half the seed, committed to compute. |
| Discovery Loop 2 days oldPublic benefit corp | Undisclosed Google, Radical, Khosla |
Jeff Dean and Sanjay Ghemawat left Google after 27 years, with Oriol Vinyals and Quoc Le, to run thousands of scientific experiments in parallel and automate the loop around them. Google is a founding investor and the cloud partner. Dean is reported as CEO. |
| Recursive SuperintelligenceSocher · Rocktäschel · ~20 people | $650M Series A Nvidia, GV, Karatage |
Raised four months after founding, before launching. Aims to automate the whole frontier pipeline — evaluation, data selection, training, post-training and research direction — with no human in the loop. Reported as oversubscribed to as much as $1B. |
| Core AutomationEx-OpenAI VP of Research | $100M @ $1B Reported raising $300–500M @ ~$4B |
Jerry Tworek led OpenAI's reasoning-model programme through 2025, then left in January. Building "the world's most automated AI lab." Flagship project Ceres targets a 100× reduction in training data through continual learning — an explicit rejection of scaling pretraining. |
| PoetiqRecursive self-improvement | $45.8M seed |
Smallest bet in the cluster and the least publicly documented. Included because the thesis is identical to Recursive's at a fourteenth of the price — a useful control on whether the capital is buying anything. |
Problem 02 · World models
Can a model predict consequences instead of tokens?
The largest cluster and the best funded. A language model predicts the next symbol. A world model predicts the next state — what actually happens if you do the thing. The disagreement inside this row is about which world: physical, visual, or economic.
Why it is still open: nobody has shown a learned world model that stays accurate over long horizons, across domains it was not trained on, without drifting into nonsense. The demos are short. The claims are long.
| Company | Capital | The specific bet |
|---|---|---|
| Skyfall AI Jul launchMaluuba founding team · Toronto | Undisclosed Fidelity, Inovia, Touring, M13, NextView, Garage |
The only team here modelling a business rather than a physical space: Enterprise World Models that simulate how a decision on pricing, hiring, marketing, finance or operations changes a company over time. Sam Pasupalak, Kaheer Suleman and Sumit Pasupalak previously built Maluuba, sold to Microsoft in 2017. First experiment: acquire a small B2B SaaS or e-commerce business for up to $1M and run it almost entirely with AI — a falsifiable public test, not a benchmark. |
| Ineffable IntelligenceDavid Silver · London | $1.1B seed Sequoia + Lightspeed co-led; Nvidia, DST, Index, Google, UK fund |
Largest seed round in European history. Silver led AlphaGo and AlphaZero; the thesis is the same one — a "superlearner" that discovers knowledge through reinforcement learning rather than by reading human data. No product, no revenue, no public roadmap. First model benchmarks were promised for late 2026, which makes this the single most checkable claim on the board. |
| AMI LabsYann LeCun · Paris | $1.03B seed Nvidia, Temasek, SoftBank VF3, Accel |
LeCun spent a decade arguing publicly that token prediction is a dead end and left Meta to fund the alternative: systems that build abstract representations of the world, hold persistent memory, and plan. Held the European seed record for 48 days. Valuation is reported inconsistently — see the conflicts log. |
| General IntuitionSister company of Medal | $320M Series A $454M total · Khosla and others |
The most contrarian data strategy in AI: train on billions of action-labelled video-game clips from Medal's 17M monthly users, on the theory that gameplay is the largest existing corpus of intent-plus-consequence. Reports say one model plays a shooter for 100 hours and then steers a quadruped robot after eight minutes of real-world fine-tuning. |
| OdysseyEx-Voyage / ex-Wayve | $310M Series B Natural Capital led; Amazon, AMD Ventures, GV, EQT, IQT |
World models for how people, objects and environments interact — targeted at robotics, autonomy, science and games. Cap table reads as an industry consensus bet: Jeff Dean, Elad Gil, Garry Tan and Kyle Vogt are all existing backers. AWS is now preferred cloud. |
| World LabsFei-Fei Li | $1.0B round, Feb a16z, Nvidia, Autodesk |
The oldest company on this board and the only one with a commercial product — Marble launched commercially the same month it raised. Thesis: the next capability gap is reasoning in three dimensions. Useful as the control case: what a world-model company looks like two years after the seed round. |
Problem 03 · Learning without the internet
Can a system keep learning after training stops?
Today's models are frozen at the end of pretraining and the internet has been read. Both facts point at the same missing capability: an agent that learns continually from its own experience, at low cost, in domains where no dataset exists.
Why it is still open: continual learning collapses. Networks trained on a stream forget what came before, and nobody has a stable, efficient algorithm at scale. It is the oldest unsolved problem here — and the one with the smallest cheques.
| Company | Capital | The specific bet |
|---|---|---|
| Oak LabRichard Sutton · Khurram Javed | Undisclosed Left Keen Technologies |
The 2024 Turing laureate and father of reinforcement learning, publicly calling generative AI a dead end at 69. Building his OaK — Options and Knowledge — architecture, and has already published a continual-learning algorithm, NetworkIDBD. The stated target: a trillion-parameter agent that learns and plans in real time on 20 watts. Smallest visible capital, largest claim. |
| Confluence TechnologiesYC W26 · Brent Burdick | YC W26 |
Works on "learning efficiency" — solving problems in domains where training data is scarce. Proof of concept: 97.9% on ARC-AGI-2 at roughly $12 per task, on a benchmark explicitly designed to be resistant to scale. Sat in the same YC batch as the benchmark's author and the foundation that runs it. |
| Core Automation (Ceres)Cross-listed from Problem 01 | $100M @ $1B |
The Ceres project is a continual-learning bet — 100× less training data — wrapped inside an automated-research company. Two unsolved problems, one cap table. |
| Ineffable IntelligenceCross-listed from Problem 02 | $1.1B seed |
"Learns without human data" is a continual-learning claim as much as a world-model claim. If the late-2026 benchmarks land, this cluster reprices. |
Problem 04 · Post-transformer architectures
Is there a different shape for the model?
Everything in production is a transformer trained autoregressively. This row is funded on the premise that the architecture, not the scale, is the ceiling — and that a different shape gets more intelligence per FLOP.
Why it is still open: alternatives keep winning on small benchmarks and losing at scale, because the transformer has a decade of hardware, tooling and optimiser tuning behind it. Beating it requires beating that entire stack, not just the maths.
| Company | Capital | The specific bet |
|---|---|---|
| NdeaChollet + Knoop · YC W26 · 15 people | $43M |
François Chollet wrote Keras and designed ARC-AGI, the benchmark the whole field uses to argue about generalisation. His bet: program synthesis fused with deep learning, not larger transformers. Mike Knoop co-founded Zapier. The smallest capital-to-credibility ratio on the board. |
| humans&Ex-OpenAI, DeepMind, Stanford, MIT | $480M seed SV Angel + Georges Harik led; Nvidia, Bezos, GV |
One of the largest seeds in tech history. Founders Andi Peng, Georges Harik, Eric Zelikman, Yuchen He and Noah Goodman are building around multi-agent reinforcement learning, memory and long-horizon planning — a "human-centric" lab whose architectural argument is that intelligence is social, and single-agent training is the wrong frame. |
| Lamb LabsYC S26 | YC S26 |
Custom inference silicon plus a method for converting autoregressive models into diffusion architectures — attacking the shape and the substrate at once. Earliest-stage entry on the board. |
| Inception LabsDiffusion language models | Undisclosed |
Generates text by denoising in parallel rather than one token at a time. The most established alternative-decoding company, and the reference point for whether diffusion LMs hold up outside a demo. |
| Logical IntelligenceEnergy-based models | Undisclosed |
Energy-based models for safety-critical systems — the architecture LeCun has advocated for years, aimed at domains where a confident wrong answer is unacceptable. |
Problem 05 · The machine that runs the experiment
Can AI do science, not just read it?
The best-capitalised cluster after world models, and the one with the clearest customer. The distinction that matters: reading the literature is solved, proposing hypotheses is nearly solved, and running the experiment and knowing what it means is not.
Why it is still open: science is bottlenecked on physical throughput and on judgement about which of a thousand results is real. Automating the pipette is engineering. Automating the taste is not.
| Company | Capital | The specific bet |
|---|---|---|
| Isomorphic LabsDeepMind spinout | $2B+ Series B |
The AlphaFold franchise turned into a drug-discovery company. The most credible route from a model to a molecule anyone has, and the benchmark the rest of this cluster is measured against. |
| Lila Sciences"Scientific superintelligence" | $13B valuation |
Autonomous labs where the model designs and physically executes experiments. Highest valuation in the cluster by a wide margin, on the least externally verified results. |
| Periodic Labs"AI scientist" | $500M Series A |
Materials and physical sciences, with a founding team drawn from frontier-lab research rather than from biotech — a deliberate bet that the bottleneck is model capability, not domain access. |
| CuspAIMaterials discovery · Cambridge | $450M Series B Lightspeed, Samsung |
Search over the space of possible materials, with corporate money attached to specific end markets — the most commercially grounded entry here, and the largest European AI-for-science round of the year. |
| Chai DiscoveryMolecular structure prediction | $400M Series C |
Open-weight structure models used widely enough that adoption, rather than valuation, is the evidence. Repriced fast in July. |
| Helical · Aemon AI · TerranoxSeed and YC stage | $10M seed · YC W26 |
The small end of the same thesis: Helical is a "virtual AI lab for biology"; Aemon AI tests hypotheses at scale; Terranox hunts uranium deposits in North America. Useful as a read on what the cluster looks like without a nine-figure cheque. |
What the board says
The money has moved from applications to premises. In 2024 and 2025 the large private rounds went to companies with revenue. In 2026 the four largest new-lab rounds on this board — Ineffable, AMI, Recursive, humans& — total roughly $3.3B raised against zero shipped product. Investors are no longer buying traction; they are buying a specific disagreement with the transformer.
Three of the five clusters are the same argument. Self-improvement, continual learning and post-transformer architectures all reduce to one claim: pretraining on a fixed corpus has a ceiling. World models and AI-for-science are the two clusters that would still matter if that claim turns out to be wrong.
02 · The money
Every record month in 2026 is one cheque
192 rounds, $394B, 1 January to 7 August. Sort it, filter it, take it away as CSV. But the finding sits above the table: strip the single largest round out of each month and the underlying US venture market has been almost perfectly flat since February.
192
Rounds in the ledger
$394B
Tracked capital
88.2%
Of AI capital to US companies
~$17B
US monthly run rate, ex-mega-round
US venture funding by month, 2026 — with and without the single largest round
$ billions. Source: AlleyWatch monthly US reports, data via Crunchbase.
Headline total
Excluding the month's single largest round
January's $6.40B does not include xAI's $20B Series E, which closed on 6 January and is counted in other datasets. That single discrepancy is the clearest illustration of why the totals on this page carry a source label: two reputable trackers, both drawing on Crunchbase, disagree by more than 3× on the same month.
The ledger
192 rounds, roughly $100M and up, January through 7 August 2026, plus a handful of smaller frontier bets. Sort by any column. Filter by month, region, sector, or to AI-only and frontier-only. Amounts in non-USD currencies are converted at approximately 7.1 CNY/USD and 1.09 USD/EUR and shown as reported. Not exhaustive — see §04 for what the coverage does and does not include.
| Company▲ | Amount▼ | Round▲ | Date▲ | Valuation▲ | Sector▲ | Country▲ | Investors▲ |
|---|
Row colour marks the region: ● US · ● China · ● Europe · ● Asia ex-China · ● other. FB marks a company on the frontier board.
Where it went: regions
Of the $362B of AI capital in this ledger, 88.2% went to US-headquartered companies — which matches the independently reported figure for 2026 AI funding, and is the single most quoted statistic about this market. It is also the least informative one.
All tracked AI capital
$361.6B across 137 rounds
Excluding OpenAI, Anthropic, xAI and Waymo
$108.6B across 133 rounds
China
17 rounds, $22.0B, and a different shape entirely. DeepSeek's ¥50B Series A — roughly $7.0B, the largest Chinese AI round on record — was the biggest disclosed round anywhere in June. Moonshot AI raised three times in 2026 ($700M in February, $2B+ in May, $3.5B in July). Below the frontier labs, the money is overwhelmingly in embodied AI: TARS (¥3.5B seed), AI2 Robotics, Galaxy Bot, Galaxea, X Square, Robot Era, Spirit AI, Lingxin Qiaoshou, Cowarobot. Alibaba, Tencent, Baidu Ventures, HSG and IDG appear repeatedly, alongside explicit state vehicles like the National AI Industry Fund. Where the US is funding models and the power to run them, China is funding bodies.
Europe
18 rounds, $15.3B, and a barbell. At one end, three of the most-watched pre-product labs in the world — Ineffable Intelligence ($1.1B seed, London), AMI Labs ($1.03B seed, Paris) and Recursive ($650M Series A) — which between them broke the European seed record twice in 48 days. At the other, real revenue: Helsing ($1.8B, the largest European AI round of the year), NEURA Robotics ($1.4B), Nscale ($2B), Wayve ($1.2B), ElevenLabs, Legora, CuspAI, Alan, Neko Health. Almost nothing in the middle. Note who leads: Sequoia, Lightspeed, Nvidia, a16z and Founders Fund are on most of these cap tables. Europe is generating the founders; the term sheets are still American.
Who is writing the cheques
Count of rounds in this ledger each firm participated in — lead or otherwise, as reported. It is a participation count, not dollars deployed, and press coverage names leads more reliably than followers, so treat it as a floor.
Most active investors, 192 rounds
Appearances in the ledger, January–August 2026
Strategic / corporate
Venture
Sovereign or platform
Concentration on the buy side too
Six firms — Nvidia, a16z, Sequoia, Index, Lightspeed and Founders Fund — appear in 101 of 192 rounds. On the frontier board specifically, Nvidia is in seven of the twenty-one bets. If the post-transformer thesis is wrong, the losses are not distributed across the venture industry; they are concentrated in a handful of balance sheets, one of which also sells the compute.
Valuation, dilution, and the price of an unsolved problem
Round size divided by post-money valuation gives implied dilution — roughly how much of the company was sold. Run it across the ledger and a clean inverse appears: the further a company is from a product, the more of itself it sells.
Implied dilution: round size as a share of post-money valuation
Selected rounds of $200M+ with a disclosed valuation
Pre-product
Early, no revenue disclosed
Product shipping
Revenue at scale
Repricing velocity
Anthropic: $380B → $965B in 15 weeks. A 154% increase between the February Series G and the May Series H. The May round was the largest venture round in recorded history and 54% of all global venture funding that month.
The European seed record broke twice in 48 days. AMI Labs' $1.03B on 10 March, then Ineffable's $1.1B on 27 April. Both pre-product. Both world-model theses.
Baseten: $300M in January, $1.5B in June. Five times the cheque, five months apart, same company, same market.
The mega-round has become the market
60% of global venture funding in H1 — around $320B — went to rounds of $1B or more. In the US that share is 73%.
23 known billion-dollar-plus US rounds through mid-2026, already level with all of 2025, with five months to go. Two rounds — OpenAI's and Anthropic's — are more than half the total US mega-round capital.
Only two of them were seed or early stage: Project Prometheus and World Labs. Everything else at that size is late-stage or corporate. The billion-dollar seed is still an exception, even now.
Handle with care
Round sizes in 2026 are unusually hard to pin down. Companies raise in tranches, announce at signing rather than closing, and count credit facilities alongside equity. Anthropic appears in this ledger in February ($30B), April ($15B) and May ($50B) — three separately reported events that do not reconcile with the aggregate "total raised" figures in the press. Valar Atomics appears twice with two different Series B rounds. Every figure here is press-reported, not audited, and the disagreements are listed in §05.
03 · Y Combinator
The batch that stopped writing software
Y Combinator ran three batches in 2026. The Winter cohort is being called the strongest in the accelerator's history, and it is also the least software-like: one in eight companies builds something physical, and the largest growth category is industrials and defence.
199
Companies in W26
1 in 8
Building physical hardware
35 → 17
Industrials/defence, vs prior batch
3×
More companies at $1M ARR than W25
YC W26 by category
199 companies across 15 categories. Shown: the categories that define the batch.
W26 · Winter 2026 — Demo Day 26 March
Rebel Fund's model scored 35% of W26 in its top-20% band — by its own account "far more than any other YC batch." Founders skew younger and more recently graduated, and the batch is more Bay Area-concentrated and far less consumer than historical norms.
The reasoning cluster
The benchmark, its author, and the company that beat it — in one batch
A genuinely unusual coincidence: ARC-AGI's creator, the foundation that runs it, and the lab that posted 97.9% on ARC-AGI-2 all sat in W26 together.
| Company | Signal | What it is |
|---|---|---|
| NdeaChollet + Knoop | $43M · 15 people | Program synthesis fused with deep learning as the path to AGI. Chollet wrote Keras and designed ARC-AGI; Knoop co-founded Zapier. On the frontier board. |
| Confluence TechnologiesBrent Burdick | 97.9% ARC-AGI-2 | At roughly $12 per task. Focused on learning efficiency in data-scarce domains. On the frontier board. |
| ARC Prize FoundationBenchmark operator | Used by OpenAI, Anthropic, Google | Runs the benchmark the frontier labs use to argue about generalisation. |
| RubricPost-training research | Exabyte scale | Post-training research infrastructure. |
The rest of W26, by what it is actually attacking
- Agent reliability — Polymath (long-running training environments), Salus (validates agent actions before execution), Arga Labs (sandboxed replicas of real services for testing), Compresr (context compression), Moda (monitoring and debugging autonomous systems), Corelayer and IncidentFox (AI on-call engineers), The Token Company (LLM input compression).
- Physical AI and the data problem behind it — One Robot (world-model simulation of gripping and assembly), Asimov (collects human-movement video to train humanoids), RoboDock (autonomous fleet charging), Servo7 and Hetherington Robotics.
- Science and resources — Aemon AI (tests hypotheses at scale), Terranox AI (uranium deposit discovery), Strand AI (biological measurement infrastructure), Axion (satellite imagery).
- Energy, because the models need it — Squid (AI grid planning), Voxel Energy (power for data centres), Condor Energy (procurement).
- AI-native services rather than software — General Legal and Moritz positioned as AI-native law firms, not vendors selling to firms. The most interesting structural bet in the batch.
- Consumer and devices — Button Computer (voice-controlled wearable, ex-Apple founders), Pocket (conversation capture), Fort (strength-training wearable), Doomersion (language learning through short video).
- Security and trust — Hex (continuous vulnerability probing), Crosslayer Labs (website spoof detection), Beesafe AI (impersonation scams), Veriad AI (marketing compliance), MouseCat (cloud-storage fraud detection).
P26 · Spring 2026
- Tasklet — a cloud agent that operates across business applications.
- Superset — open-source IDE for coordinating parallel coding agents.
- Arga Labs — sandboxed service replicas for agent testing.
- Silmaril — runtime protection against prompt injection.
- Complir — compliance for physical products.
- 9 Mothers — autonomous counter-drone systems.
- Ploy — agents that optimise websites and campaigns.
S26 · Summer 2026 — Demo Day 10 September, running now
Incomplete by definition: the batch runs July through September and has not pitched yet. Early names worth a watch:
- Lamb Labs — custom inference silicon, plus converting autoregressive models into diffusion architectures. On the frontier board.
- Tracer — a research lab studying coordinated intelligence across AI systems.
- Experiential Labs — world models to improve agent understanding.
- Grip Robotics — physical AI grippers for waste and materials handling.
- Decawork — an agent control plane for IT: deploy, govern and maintain agents company-wide.
- Justinian — pitched as the first AI government-affairs firm.
- Conifer — model routing that balances cost against data locality.
- Instance — automated labelling for robot training and evaluation.
- Context.dev — real-time structured web data as an API for agents.
What YC 2026 tells you that the funding data does not
The seed layer has stopped competing with the labs and started supplying them. The mega-rounds in §02 go to companies training models. The batches in §03 mostly go to companies making those models safe to deploy, feeding them data, or wiring them to physical systems. Only a handful — Ndea, Confluence, Lamb Labs — take the labs on directly, and all three do it by rejecting scale rather than trying to match it.
04 · How to track this
The system, not the list
A list of frontier labs is out of date the week you publish it — three of the companies in §01 did not exist in April. What survives is the method. This is the one behind this page: where to look, what a pre-launch lab looks like before it launches, and the errors that make most AI funding data wrong.
The source stack
Four tiers, used for different things. The mistake is treating them as interchangeable — the aggregators are for discovery, the business press is for facts, and confusing the two is how a valuation becomes a round size.
| Tier | Use it for | Sources |
|---|---|---|
| 1 · Primary pressFacts of record | Amounts, valuations, investor lists, founder names. The only tier you should quote a number from. | TechCrunch, Crunchbase News, CNBC, Forbes, The Information, Bloomberg. Plus the national outlets for non-US rounds — The Logic in Canada, SiliconCanals and Sifted in Europe. |
| 2 · Monthly roundupsSystematic coverage | The backbone of a ledger. One source, one methodology, every month — which is what makes month-over-month comparison legitimate. Everything in §02 that is charted comes from here. | AlleyWatch's monthly "largest global rounds" and "US venture capital funding report" (data via Crunchbase), Crunchbase News monthly recaps and weekly "10 biggest rounds", Intellizence monthly. |
| 3 · Specialist trackersDiscovery only | Finding companies you have never heard of. Fast, wide, and wrong often enough that nothing should leave this tier without being re-checked against tier 1. | The neolab and emerging-lab trackers, AI funding aggregators, Dealroom, Tracxn, PitchBook profiles. Every candidate on this page's frontier board was found here and verified elsewhere. |
| 4 · Primary sourcesWhat they actually claim | The company's own words, which are usually more specific than the coverage. A careers page tells you the size and shape of the team before any journalist does. | Company sites and manifestos, arXiv and lab publication pages, ycombinator.com company directory, founder posts on X and LinkedIn, cloud-partnership announcements. |
What a frontier lab looks like before it launches
Every company in §01 was visible for weeks before it was news. These are the tells, in rough order of how early they appear.
- A named researcher leaves a frontier lab without saying where they are going. Jerry Tworek left OpenAI in January; Core Automation launched in May. Two Anthropic researchers left; Mirendil raised $200M four weeks later. This is the earliest signal available and it is public.
- A one-page site with a thesis and a careers link, and no product page. Mirendil's site describes a problem and a hiring pipeline. That is the shape of a lab that has money and no product yet.
- A cloud or compute deal announced before anything else. Compute commitments leak before launches because they involve a second company with its own comms schedule. Mirendil's $100M+ Google Cloud deal, Odyssey's AWS agreement, World Labs' Autodesk deal.
- A benchmark result posted without a company attached to it. Confluence's 97.9% on ARC-AGI-2 circulated before most people knew the company's name.
- Nvidia on the cap table of something with no revenue. It appears in 26 rounds in this ledger and is the most reliable single marker of a company that intends to spend heavily on training.
- An unusual acqui-target or public experiment. Skyfall announcing it will buy a company for $1M and run it with AI is not a product launch — it is a falsifiable claim, which is rarer and more informative.
The weekly routine
Twenty minutes, once a week, keeps a tracker current. In order:
- Read one weekly roundup end to end. Crunchbase News' "The Week's 10 Biggest Funding Rounds" is the single highest-yield item. It takes four minutes and catches most things over $100M.
- Scan for the three words that matter — "emerges from stealth", "co-founded by", "left to start". Everything on the frontier board was announced with one of those phrases.
- Check the departures. Who left OpenAI, Anthropic, Google DeepMind or Meta this week, and did they say why. This is the leading indicator; the funding round is the lagging one.
- Run the five tests from §01 on anything new. Four out of five gets a row. Fewer than four is an application company, which is a different tracker.
- Record it, and record the source. One row, with the URL. A number without a source is not data, it is a rumour you will believe in six weeks.
- Once a month, reconcile. Pull the monthly roundup and diff it against what you captured weekly. What you missed tells you which source to add.
Queries worth saving
"emerges from stealth" AI lab # launches
"left OpenAI" OR "left Anthropic" OR "left DeepMind" to found
site:ycombinator.com/companies <batch> AI
"largest seed" OR "record seed" AI <year>
"world model" OR "continual learning" OR "self-improving" raise
<company> "Series" -site:pitchbook.com -site:tracxn.com # skip paywalled aggregatorsThe four errors that make most AI funding data wrong
ERROR 01Valuation read as round sizeThe most common error in the field. General Intuition is widely listed as a "$2.3B Series B." It was a $320M Series A at a $2.3B valuation — a 7× overstatement. Always check whether the big number is what went in or what it was worth after.
ERROR 02Tranche double-countingLarge rounds close in stages and get announced more than once. OpenAI's $122B had a $110B first tranche. Anthropic shows up three times in 2026. Prometheus appears as $10B in April and $12B in June. Aggregate "total raised" figures rarely reconcile with the monthly events.
ERROR 03Debt counted as equityValar Atomics' $1.2B "total raised" includes a credit facility. Data-centre and energy companies do this constantly, because the capital structure genuinely is part debt. An equity comparison that includes debt is not a comparison.
ERROR 04Dataset driftTwo trackers both citing Crunchbase reported US January 2026 at $6.40B while xAI's $20B closed on 6 January. Different inclusion rules, same underlying database. Never mix sources inside one chart, and always label which one you used.
How this page is maintained
The ledger is not typed into the HTML. data/rounds.json is the single source of truth; a build step regenerates the CSV export and rewrites the inline data block in the page. Updating the tracker is three commands:
# 1. add or edit rows in the dataset
$EDITOR data/rounds.json
# 2. regenerate data/rounds.csv and the inline block in index.html
python3 build.py
# 3. publish
git commit -am "ledger: <what changed>" && git pushEach row carries company, amount_usd_m, amount (as reported), round, date, valuation, sector, country, region, ai, frontier, investors and a free-text note for conflicts. The note field is the important one: it is where a disagreement between sources gets written down instead of silently resolved.
The one rule
Record the disagreement rather than picking a winner. Nine of the figures on this page are reported differently by two credible sources. A tracker that silently picks one and moves on looks more authoritative and is less useful — because the disagreements are themselves a signal about which numbers are soft. That is what the conflicts log in §05 is for.
05 · Watchlist & sources
What would change this page
A tracker is only useful if it tells you what to look for next. These are the dated, checkable events between now and the end of 2026 that would move companies on or off the frontier board.
LATE 2026
Ineffable's first benchmarks
Silver's lab said first model benchmarks would arrive by late 2026. $1.1B was raised against a claim with a date on it — the cleanest test on this board.
WATCH: any published eval
OPEN
Skyfall's acquisition closing
The plan is to buy a small B2B SaaS or e-commerce business for up to $1M and run it almost entirely with AI. Which company, and what the P&L does afterwards, is a real-world result — not a demo.
WATCH: named target, then quarterly numbers
Q3–Q4 2026
Core Automation's round
Reported to be raising $300–500M at ~$4B, weeks after closing $100M at $1B. If it prices, automated-research labs have gone 4× in a quarter on no product.
WATCH: confirmed close and valuation
NEXT 90 DAYS
Discovery Loop's raise and hires
Announced 5 August with no disclosed amount. Given the founders, the size of the first round and who else leaves Google to join will set the price for the whole self-improvement cluster.
WATCH: round size, departures from Google Research
OPEN
Oak Lab's funding
No disclosed capital, a Turing laureate, and the most aggressive claim on the board — real-time learning at 20 watts. Whether serious money attaches to Sutton's architecture is the market's verdict on continual learning.
WATCH: any announced round
10 SEP 2026
YC S26 Demo Day
The first cohort assembled entirely after the 2026 mega-rounds. Whether it keeps W26's deep-tech and physical tilt, or reverts to application software, is the leading indicator for 2027 seed.
WATCH: physical-AI share, foundational-research count
ONGOING
Whether ARC-AGI-2 stays beaten
Confluence posted 97.9% at ~$12 per task. If that reproduces independently — and if ARC-AGI-3 resets it — the case for scale-rejecting architectures gets its first hard evidence.
WATCH: independent replication, ARC-AGI-3 results
Q3 2026
Whether the underlying market moves
Ex-mega-round US funding has sat at roughly $17B a month since February. If that line finally breaks upward, the boom is broadening. If a mega-round month arrives and the line stays flat, it is not.
WATCH: the dashed line in §02
Conflicts log
Every figure below is reported differently by at least two credible sources. Where this page had to choose, it took the primary business-press number over aggregator databases, and said so.
- US January 2026 total. Reported at $6.40B by AlleyWatch (data via Crunchbase) in a month when xAI's $20B Series E closed on 6 January. Different inclusion rules on the same database. Page shows $6.40B, labelled, and flags the discrepancy on the chart.
- General Intuition. Listed by trackers as a "$2.3B Series B." Primary reporting is a $320M Series A at a $2.3B valuation. Page uses the primary figure. This is the most common error class in AI funding data.
- Anthropic's 2026 raises. $30B (Feb), $15B (Apr) and $50B (May) are separately reported. Some aggregators report a single "$132B Series H at $965B." These do not reconcile; the page lists the three monthly events.
- Project Prometheus. Crunchbase reports $10B in April; other trackers report a $12B Series B at a $41B valuation dated June. Tranches of one raise, or two events — unresolved. Both rows are kept so the disagreement stays visible, which means the ledger totals may double-count up to $12B — about 3% of the tracked capital. The June row is labelled Series B (disputed).
- Valar Atomics. A $340M Series B in April and a $1.0B Series B at $6B on 4 August. Total raised is given as $1.2B including a credit facility, which does not reconcile with $340M + $1.0B in equity. Both rows are in the ledger.
- Recursive. Reported as ~$500M pre-Series A at a $4B pre-money, US-based (The Decoder, MLQ) and as a $650M Series A, UK-based (AlleyWatch). Page uses $650M and notes the earlier figure.
- Shield AI. $2.0B at up to $18B (AlleyWatch, March) versus $1.5B at $12.7B (Crescendo, 25 March). Page uses $2.0B.
- Fundamental. $225M with Valor and Headline (AlleyWatch) versus $255M with Oak HC/FT and Salesforce Ventures (TechCrunch). Both amount and leads differ. Page uses $225M and lists all four investors.
- AMI Labs valuation. The $1.03B seed is consistent; the valuation is reported at $3.5B in some accounts and $5B in others. Page shows the range.
- H1 2026 total. $510B (Crunchbase H1 report) versus $515B (Crunchbase July report, Jan–Jun). Page uses $510B.
- Q1 2026 total. Reported as $297B, $300B and $305B across three outlets. Page uses $300B.
- Maluuba's sale price. US$140M (Wikipedia) and ~US$160M (2026 founder profiles). Terms were never disclosed. Relevant as the Skyfall founders' track record.
- Undisclosed by design. Skyfall, Discovery Loop, Oak Lab, Inception Labs and Logical Intelligence have no public round size. Listed on thesis and team, not capital.
- Thin sourcing, flagged not dropped. Poetiq, LMArena and Flapping Airplanes rest on aggregator entries or a single roundup.
Method
- Scope. Compiled 7 August 2026. Ledger covers 1 January – 7 August 2026. August is a partial month.
- Ledger coverage. 192 rounds of roughly $100M and up, plus a small number of smaller frontier bets, assembled from published monthly and weekly roundups. Not exhaustive — a round that no roundup covered is not here, and coverage is thinner for January and for non-US rounds below $200M. Treat totals as a floor, not a census. The one exception runs the other way: Project Prometheus appears twice (April $10B, June $12B) because the sources disagree on whether that is one raise or two, so totals may overstate by up to $12B.
- Inclusion on the frontier board. A company qualifies if its stated thesis targets a capability that no shipped system currently has, and if it passes at least four of the five tests in §01. Companies with revenue from an existing product are included only where they are the reference case for a cluster (World Labs, Isomorphic).
- Charts. The monthly US series is AlleyWatch's US venture capital funding reports, data via Crunchbase — one source, one methodology, so the months are comparable. The "excluding the largest round" figures are published directly for February and May and derived by subtraction for March, April, June and July.
- Currency. Non-USD rounds are shown as reported with an approximate USD conversion at roughly 7.1 CNY/USD and 1.09 USD/EUR, used only for sorting and totals. Treat as indicative.
- Investor counts are participation counts from reported investor lists, not dollars deployed, and press coverage names leads more reliably than followers. Read them as a floor.
- Not verified. Valuations, round sizes, headcounts and benchmark scores are press-reported and not independently confirmed. Benchmark claims — including Confluence's ARC-AGI-2 result — are self-reported unless stated otherwise.
- Editorial. Cluster assignments, the five tests, the four error types and the watchlist are judgement, not fact. Reasonable people would sort several of these companies differently.
- Independence. Not affiliated with, endorsed by, or published by any company named. No sponsorship, no tracking, no analytics.
Principal sources
01TechCrunch — "Jeff Dean and other top AI researchers are leaving Google to launch their own startup," 5 Aug 2026
02TechCrunch — "Exclusive: Mirendil inks $100M+ Google Cloud deal to scale self-improving AI," 6 Aug 2026
03TechCrunch — "DeepMind's David Silver just raised $1.1B to build an AI that learns without human data," 27 Apr 2026
04TechCrunch — "General Intuition's $2.3B bet that video games can train AI agents for the real world," 25 Jun 2026
05TechCrunch — "What happens when AI starts building itself," 14 May 2026
06TechCrunch — "Here are the 17 US-based AI companies that have raised $100M or more in 2026," 17 Feb 2026
07TechCrunch — "16 of the most interesting startups from YC W26 Demo Day," 26 Mar 2026
08AlleyWatch — "The 23 Largest Global Startup Funding Rounds of February 2026"
09AlleyWatch — "The 16 Largest Global Startup Funding Rounds of March 2026"
10AlleyWatch — "The 17 Largest Global Startup Funding Rounds of April 2026"
11AlleyWatch — "The 22 Largest US Funding Rounds of May 2026"
12AlleyWatch — "The 19 Largest Global Startup Funding Rounds of June 2026"
13AlleyWatch — US Venture Capital Funding Reports, March / April / May / June / July 2026
14Crunchbase News — "Global Startup Investment Hit Record $510B In H1 2026"
15Crunchbase News — "Q1 2026 Shatters Venture Funding Records"
16Crunchbase News — "Anthropic Funding Pushed Startup Investment To Near-Record Levels In May"
17Crunchbase News — "Billion-Dollar AI Rounds Push April To Third-Highest Startup Funding Month In A Year"
18Crunchbase News — "A Record 14 Billion-Dollar Rounds In July Pushed Venture's Historic Run Higher"
19Crunchbase News — "The Rise And Rise Of Billion-Dollar-Plus Rounds," H1 2026
20Crunchbase News — weekly "The Week's 10 Biggest Funding Rounds," June–August 2026
21Intellizence — "Startup Funding Trends in July 2026"
22Crescendo AI — running list of 2026 AI VC investment deals
23CNBC — "Ex-DeepMind David Silver raises $1.1 billion for AI startup Ineffable," 27 Apr 2026
24Forbes — "Former Microsoft AI Leaders Are Spending $1M To Prove AI Can Replace CEOs," 20 Jul 2026
25Forbes — "21 most promising startups from Y Combinator's latest batch," 16 Mar 2026
26Tech Startups — Skyfall AI out of stealth (20 Jul 2026); humans& $480M seed (20 Jan 2026); VC roundup 4 Aug 2026
27The Logic — "Skyfall AI wants to buy a small tech business and run it with AI," 2026
28The Decoder — Core Automation launch; Recursive Superintelligence $500M, 2026
29The Next Web — "Richard Sutton leaves Carmack to start his own AI lab"; "Ex-Anthropic researchers raise $200M for self-improving AI," 2026
30SiliconANGLE — "Mirendil raises $200M to speed up scientific research with AI," 25 Jun 2026
31HPCwire / AIwire — Odyssey $310M Series B; Unreasonable Labs out of stealth, 10 Mar 2026
32CB Insights — "Y Combinator's Winter 2026 batch is its most technically complex cohort yet"
33Jared Heyman / Rebel Fund — "On the freakishly strong YC W26 batch"
34StartupHub.ai — Claude's Corner profiles of Ndea and Confluence Labs, YC W26
35cleverhack — Neolab and Emerging AI Lab Tracker (candidate discovery; every entry re-verified against primary reporting)
36Company sites — mirendil.com, Oak Lab publications (OaK architecture, NetworkIDBD), ycombinator.com company directory
The data
The ledger is published as structured data alongside this page. Both are generated from the same source file, so they cannot drift apart.