techstartups @techstartups

Island raised $400M at a $6.4B valuation. More than double the $3B round in 2024. Last year’s secondaries cleared at $5B. Evolution Equity led. Sequoia, Coatue, Insight, Cyberstarts, and J.P. Morgan all came back in. The round is not about a browser. Island shipped a Chromium build enterprises can actually govern. Clipboard, downloads, screenshots, whatever an employee pastes into ChatGPT. Work moved into the browser, so the security stack had to follow. Same layer, new user: AI agents. An agent with access to your SaaS apps is an employee you cannot watch. CEO Mike Fey, ex-Symantec president: every old control is breaking. That’s the fight. Palo Alto already bought the other enterprise browser, Talon, and folded it into Prisma Access. You get it if you already live in their SASE stack. LayerX and Seraphic don’t ask you to switch browsers. They ship an extension on the Chrome you already have. Island’s bet is the harder sale. Rip out Chrome. Own the session. They’ve doubled ARR every year since launching in 2022 doing it. $6.4B says the browser stops being a feature inside someone else’s suite.

Snorkel AI raised $350M at $3.5B today. Skip the headline. The round prices them at about 9x revenue. $375M ARR, up 18x in a year. Insight Partners and S32 co-led. Greylock, Lightspeed, GV, and Addition were already in. Seventeen months ago this was a $1.3B company. The product is the only part that makes the multiple make sense. Snorkel came out of the Stanford AI Lab selling labeling software. Last year they switched to data-as-a-service: finished datasets, built by their models plus domain experts. Not a pure marketplace of humans. That puts them next to three companies that are not the same business. Scale lost the labs when Meta took 49% and Alexandr Wang left. Neutrality was the product, and it broke. Mercor took the expert-marketplace seat. Roughly $10B now, about $2B in gross revenue. That gross figure is throughput, not software ARR. Surge took the frontier-lab work Scale used to own. Handshake and Micro1 are scaling the same expert-hour model. So Insight and S32 just paid a software multiple for a company standing in a staffing market. If the experts are the moat, 9x is rich. If the software is actually producing the data, this is the cheapest serious asset left in AI training data. The growth doesn’t settle it. The multiple is the bet

Mach Industries just raised another $600M and doubled its valuation to $3.7B. In three months. The founder is 22. Ethan Thornton dropped out of MIT at 19. This was Sequoia’s first defense bet. June: $300M at $1.8B. Today: $600M more, same names, twice the price. Ribbit, Infinite Capital, Bedrock, and Sequoia all doubled down. Total raised now sits over $1B. Here’s the number most people will skip. Mach’s lifetime prime contracts in federal databases total under $2 million. Anduril is at $61B with real revenue. Shield AI is at $12.7B. Helsing just took $1.8B. Mach is pricing like it already won that race. The bet isn’t the drone. It’s the factory. Viper is a one-way attack UAS. Atlas launches off ships. The actual product is a Huntington Beach line that builds airframes, jet engines, and rocket motors in-house, because the bottleneck in a real war is not software. It’s who can stamp out the next thousand units. Anduril owns the software stack. Palantir owns the data. The primes spent 40 years letting the industrial base atrophy. Mach is trying to own what they abandoned. A 22-year-old just raised a billion dollars to become a defense prime. The scary part for Lockheed isn’t the pitch deck. It’s the production line. Investors: Ribbit, Infinite Capital, Bedrock, Sequoia (Stephanie Zhan, Shaun Maguire). Earlier: Khosla, RDC. Competitors: Anduril, Shield AI, Helsing, Castelion, General Atomics, Lockheed.

Crusoe just closed $3.9B at a $30.9B valuation. Eight years ago it was a crypto miner on flared gas in North Dakota. Today it builds OpenAI’s Stargate campus and sells AI factories you can drive to a spare substation on a truck. Atreides, Mubadala, and Valor led. Founders Fund, Nvidia, GIC, QIA, and TPG followed. Three sovereign wealth funds and the company that makes the chips, all writing checks into the same bottleneck: power. GPUs are not the scarce resource anymore. Megawatts are. CoreWeave leases other people’s buildings. Lambda rents you a cluster. Nebius and Fluidstack are in the same race. Crusoe went the other direction — generate the electricity, build the campus, then sell the tokens. CEO Chase Lochmiller calls it “electrons to tokens.” $140B in contracted value. 6GW of capacity. 1GW already live. The companies that win AI won’t have the best model. They’ll have found a spare gigawatt first.

Oracle's $300 Billion bet on OpenAI is starting to crack. About $18B of loans tied to a data centre leased to Oracle in New Mexico just slid into stressed territory. The banks can't find buyers for the debt, & phase one is already 7+ months behind schedule. OpenAI has deferred its IPO & projects a $280B cash burn by 2030. Last week Larry Ellison abruptly cancelled a $7.5B share plan, no explanation given. Draw your own conclusions....

Google needs to buy Instinct, they need that product DNA, Notebook isn't going to cut it. Instinct, Muse, Grok Bot, etc., are social ai. I've seen this play out before... I was at Google+ for 3-months running mobile, I tried buy instagram but I was two steps down from Larry, so I didn't have the pull (Kevin had told me Zuck was sniffing around), my SVP said it was 'too expensive.'.. killed me. I bounced to Google Ventures.

Update on my bet that OpenAI has a 50/50 chance of running out of money by next summer. I know I may be wrong here— this is what 50/50 means. But here’s how I see it. The positive: OpenAI has overtaken Anthropic in terms of weekly revenue, according to OpenRouter data from July. Reflects OpenAI’s scientific excellence, which has always stood in contrast to its shaky business foundations. Impressive that @sama and @gdb have lost loads of senior scientists and yet they still replenish the brain trust. Not so positive: Anthropic appears to enjoy far more customer loyalty. Twelve month retention rates are almost double OAI’s. (H/T @FT). But, tbf, this may just reflect a period when Anthropic’s models improved faster, and maybe the tables have turned. The chief OAI worry is the same as it always was: cash burn. Latest numbers indicate that OAI expects to lose $280 billion by 2030, even though revenue is forecast to 10x between now and then. That implies a massive fund-raising challenge. OAI is running an experiment not only in an unprecedented technology; it is also testing the depth of global capital markets. OAI has used the recent hacking scares to postpone its IPO, which might have been hard to pull off anyway. So it has to raise money privately. Already, the last fund-raising round included a bunch of financial engineering—conditional future cash promises, promises of compute in lieu of cash. OAI has also tried to finance data centers by leaning on Nvidia in its guise as the central bank of AI. See the $105 billion NVidia guarantee on the Ohio data center. How many more nimble financial tricks can OAI pull? Especially in a climate where the political backlash against AI is growing and open weight competition is intense? Here’s what to watch for. OAI financial partner, SoftBank, is plotting the near-term IPO of its data-center builder SB Energy. If that plan runs into trouble, there will be knock-on effects for SoftBank and hence a hit to SoftBank’s ability to pour fresh money into OAI.

No wonder OpenAI delayed their IPO plans. Secondary shares are being offered at an implied $475 bn valuation, down roughly 44% from the $852 bn primary round earlier this year, that makes that $1 tn IPO look near impossible right now. Hearing Anthropic is being offered around a $700 bn valuation on the secondary market, yet they’re still trying to pitch their IPO between $1 tn and $2 tn. Good luck with that. We all know how flaky secondary markets can be, so make of it what you will.

Temporal just raised $550M at $12.55B. Seven months ago they were worth $5B. The number that matters more is 200%. That’s net dollar retention. Customers aren’t sampling this. They’re drowning in it. OpenAI’s usage grew 60x in under a year. Snap moves 414 million Stories a day on it. JPMorgan runs it in regulated production. Revenue just crossed $250M ARR, up 200% YoY. Lightspeed led. Wellington, Goldman Sachs, and Tiger Global co-led. Sequoia, a16z, and Index came back in. That’s the entire AI infrastructure stack writing the same check. The competitors are obvious: AWS Step Functions, Airflow, Inngest, Trigger.dev. All of them orchestrate work. None of them were built for an agent that has to survive for weeks. Agents don’t die because the model is dumb. They die because the workflow crashes at step 47 and nobody saved the state. Temporal’s whole product is that crash never happened. The AI race was supposed to go to whoever trained the best model. It’s going to whoever keeps the agent alive when the model times out.

😳 Brad Gerstner on the ex-Anthropic researcher who told America AI could end humanity in three years: "To go on national television in a highly coordinated way and sit in the living room on NBC, on ABC, on CBS, on Fox, and have a single researcher who spent six weeks at Anthropic tell the moms and dads of America that this can kill all of humanity in the next three years without any countervailing conversation is deeply irresponsible. Jensen Huang yesterday called it deeply untrue. Andrew Feldman, the founder of Cerebras, who spent decades studying this, called it utter horseshit. Elon himself said it was a coordinated psyop." - @altcap with Scott Wapner @TheJudgeCNBC on CNBC Friday, Sept 11th, 2026

Next 6–12 months are going to be wild. Astra just trained on 100k+ Grace Blackwell. Jensen: AGI has arrived. 400k more GPUs coming online. H2’26–H1’27: Rubin racks start shipping. Agents that actually run the computer. Coding + science get compressed. Compute does not pause. 🔥

Jensen Huang @JensenHuang

a month ago

@ChaseLochmiller @OpenAI GPT-6 Astra, trained on ~100K+ NVIDIA Grace Blackwell NVLink72. From ChatGPT to o1 to Astra in 4 years. AGI has arrived. Congratulations @OpenAI team. 400K GPUs coming online next.

2K 5K 42K 12.2M 7K

Greg Brockman says Astra's breakthrough is computer use, a near-universal connector that reaches even the software with no API "You have some pieces of software, humans can use it just fine, the AI has no access to it." "So what do you do? You have to write a very specific connector that hooks up to the APIs, and not everything's exposed, so you can't do everything that you could." "Then you think about that there are so many pieces of software that don't have APIs, and those are totally out of bounds. So we have this limited world where the AI is so restricted from helping you." "I think that we now have the technology that's almost this universal connector."

Fireside Alpha @firesidealpha

a month ago

Greg Brockman says 300 million people come to ChatGPT with health questions every week, and reveals OpenAI is building a bottoms-up clinicians product, top-down enterprise product for hospitals, and a three-sided marketplace in health "It's actually very surprising to me how

2 5 50 23K 21

Gimlet Labs just raised $300M at a $3B valuation. a16z led. Arm, Microsoft’s M12, Samsung Ventures, Tiger Global, Menlo, and Sapphire piled in. They are not selling more H100s. They break a model into modules and park each on the chip that fits the job. Memory-heavy pieces go one place; compute-heavy pieces go another. That’s the product: inference is no longer one workload on one GPU. Fireworks, Groq, and CoreWeave still mostly sell a pile of silicon and hope the compiler saves you. Power is the real constraint. Tokens per kW is the feature. Companies that treat inference as a systems problem, not a hardware catalog, will survive the next 20x in token demand.

Fasset just hit unicorn. $68M Series C at $1B, led by SBI Group. Stablecoin neobank processing $40B+ annualized volume across 125 countries. 3M wallets. Profitable for 12 straight months. Revenue up 6x. Not another crypto wallet. They’re building the rails: AI-routed settlement across 100+ banking corridors, any asset to any asset. SBI isn’t just writing a check. They’re plugging Fasset into Japan’s financial network and Remit corridors. Key competitors: Wise on the traditional side, Circle on the stablecoin rails, and a long list of regional neo-banks that still move money the slow way. The quiet ones building actual payment infrastructure keep winning.

Rillet just became a unicorn. $100M Series C at $1B, led by ICONIQ with Sequoia and a16z returning. Third raise in 14 months. Total capital now >$200M. Two years out of stealth. 600+ customers. ARR doubled in the last three months. One of them (Mercor) is running $2B+ ARR with a three-person finance team. The product doesn’t sit on top of the general ledger. The agents live inside it. Real-time, full audit trail, human approval only when it matters. Oracle, SAP, NetSuite, and Workday spent decades building systems of record. Rillet is turning the ledger into the actual operating system for finance. Legacy ERP was built for a world where humans did the work and software stored the receipts. That world just got a serious problem.

Three Harvard dropouts. 
First-pass silicon success on TSMC. $5B valuation in December. 
$10.3B in July. 
Today: $700M raised at $21B. Led by Jane Street — who also became the first customer and put a rack into production. 
Joined by Sequoia, a16z, Kleiner Perkins, Tiger Global, Bain Capital Ventures, Blackstone, and Peter Thiel. Etched builds full frontier inference clusters — not GPUs. 
Low-voltage prefill. Cluster-scale memory for decode. 
More tokens per dollar. More tokens per watt. Over $1B in signed contracts. 
The inference machine just shipped its first unit. Nvidia’s still the king. 
But the specialized challengers just got real.

This is serious competition for NVDA. The startup says it took just 44 days after getting its test chips back from Taiwan Semiconductor Manufacturing to have them up and running inference workloads—the computing processes that allow AI models to respond to user queries—a process that usually takes six months or more. When Etched needed help designing servers and racks to integrate its chips into one big system, the company in 2024 recruited Brian Loiler, a top systems engineer who spent nearly 23 years at Nvidia. Loiler has since recruited to Etched about a dozen engineers from Nvidia who, in some cases, turned down attractive counteroffers, he said. wsj.com/tech/ai/a-21-b…

Lovable just raised $400M at a $13.3B valuation. Eight months ago they were at $6.6B. ARR is racing toward $600M. 60 million projects already shipped. Led by Menlo Ventures, co-led by EQT’s Scaleup Europe Fund. New capital from Balderton, Carmignac, Kaszek, Tencent, World Innovation Lab, Regent and others. Returning: Accel, CapitalG, DST Global, Salesforce Ventures, HubSpot Ventures. This is what happens when you delete the last gate between idea and working software. The PM who used to write PRDs now ships the product. The founder who used to wait on engineers now iterates in the same afternoon. Coding stopped being the bottleneck. Taste became the scarce resource. And the market just put a $13.3B price tag on it.

Excellent, candid back & forth on complex subjects that benefit from a good faith Saturday afternoon exchange! Thx @_sholtodouglas & @GavinSBaker for kicking off & super helpful to have @DarioAmodei weigh in. 🙏

Dario Amodei @DarioAmodei

2 months ago

1/2 Thanks Gavin for an especially thoughtful exchange. I don't usually spend much time on social media but I wanted to engage here because it really brings out the heart of an important conversation. First, on regulation, I think that “either concentrate it in the hands of a

1K 911 9K 7.6M 9K

Hadrian just raised $1.37B at nearly $8B. Not another AI chatbot. Not another agent wrapper. They’re building automated factories that mass-produce the parts Lockheed, RTX, Anduril, and the Navy actually need — submarines, munitions, drones. Founded 2021. Four sites. Almost 3 million square feet. Software called Opus runs the whole thing like a product, not a plant. While everyone else was arguing about tokens, Chris Power bet that the real bottleneck was atoms. America’s ability to make physical things at speed. This round is the market admitting he was right. Physical AI just got its biggest check of the year.