Newsletter · · Ashutosh Agarwal

Fireworks Raises 1.5 Billion as Anthropic Approaches a Trillion Dollar Valuation - The Raise - Week of Jul 20–Jul 26, 2026

Startups and venture newsletter for the week of July 20 to 26, 2026. Anthropic's $65 billion raise at a near trillion dollar valuation leads a heavy week of rounds, with Fireworks founder Lin Qiao on inference economics and the risk of scaling into bankruptcy.

The Raise

Week of Jul 20–Jul 26, 2026: Fireworks Raises 1.5 Billion as Anthropic Approaches a Trillion Dollar Valuation

Week of Jul 20–Jul 26, 2026: the week a startup was worth almost a trillion dollars, a pool table got smarter than an F-35, and an inference company nobody outside the Valley has heard of raised $1.5 billion.

On July 23, a private company that is four years old and losing money hand over fist became worth nearly a trillion dollars. Anthropic, the maker of Claude, closed a $65 billion raise (yes, that is the amount raised, not the valuation) at a valuation the podcasts pegged at "nearly $1 trillion," officially passing OpenAI as the most valuable AI startup on earth. It is the kind of number that stops the tape.

Here is the part worth chewing on. The headline number everyone repeats is Anthropic's $47 billion of annualized revenue, a figure that would make it one of the fastest-scaling businesses in history. But on the Elon Musk Podcast's breakdown of the round (2026-07-23), the hosts did the unglamorous work of asking where that money actually comes from, and the answer is messier than the headline. A big chunk of it is billed through the cloud providers, Amazon (AMZN) and Google (GOOGL), who "take a substantial cut before the cash actually hits the model developer's balance sheet." In other words, the same companies routing customers to Anthropic are also selling it the compute and keeping a toll on the way through. The top line is a gross number, not a clean one.

And underneath it, the company is bleeding: Anthropic internally projects a $14 billion loss. As the podcast put it, both AI giants are running the old rideshare playbook, subsidize the customer now, lock in the behavior, pray the cost of compute falls before the runway does:

"they are just stuck selling a dollar for 50 cents indefinitely."

So why would sophisticated investors pour $65 billion into a dollar-for-fifty-cents machine? Because this round is not really a normal financing, it is a hardware land-grab dressed up as an equity raise. About $15 billion of it came from prior hyperscaler commitments, including $5 billion from Amazon, plus capacity deals for next-generation TPUs from Google and Broadcom (AVGO), memory from Micron (MU), Samsung and SK Hynix, and GPU access inside SpaceX's Colossus data centers. Anthropic is trading equity for a guaranteed seat at the silicon, memory and power table. That tells you what the real bottleneck is in AI right now, and it is not money, it is whether you can physically plug in enough chips.

The valuation is hard to defend on any metric a normal analyst would use (roughly 20x a revenue number that is partly other people's gross spend). But the enterprise traction is not a mirage: the podcast cited Anthropic holding 54% of the coding market to OpenAI's 21%, and capturing 73% of first-time enterprise buyer spend. Developers vote with their keyboards, and right now they are voting Claude. The scary part for everyone building on top of these models is the flip side of that lock-in: if the trillion-dollar landlord ever decides to stop losing money and raise the rent, your margins are the first thing that disappears. Keep that in the back of your head every time a founder tells you their whole product is a wrapper.

Now, to the rounds.

This Week's Rounds

  • Anthropic, $65B raised at a ~$1T valuation, Series H, backed by a syndicate that includes hyperscaler capacity commitments ($5B from Amazon (AMZN), TPUs via Google (GOOGL) and Broadcom (AVGO), memory from Micron (MU), Samsung and SK Hynix, GPUs in SpaceX's Colossus). The AI lab behind Claude; ~$47B annualized revenue against a projected $14B loss. The money buys compute, not comfort.
  • Fireworks, $1.5B at a $17.5B post-money valuation, led by Index Ventures, Gavin Baker, Lightspeed and 20VC. The inference layer that runs open-source AI models for enterprises; over $1 billion in ARR in about four years, now serving ~40 trillion tokens a day, up from 15 trillion. (This is also the Founder Story below.)
  • Atoms, $1.7B led by Andreessen Horowitz, with Uber (UBER) participating. Travis Kalanick's "industrial AI and physical automation" venture, which just acquired Anthony Levandowski's autonomous-mining startup Pronto. On Equity (2026-07-24), the hosts were openly skeptical: nobody is quite sure what the company does yet, and part of the raise looks like the price of getting to post "Travis is back, baby" on Twitter for 24 hours straight.
  • CuspAI, $450M (on top of a $100M Series A last year). AI for materials discovery, meaning generative models that invent new molecules and materials, grounded in a global network of physical labs, with a big NVIDIA (NVDA) partnership. Founder Chad Edwards has done something genuinely novel: baked royalties into customer contracts, so CuspAI shares in the upside when a material it discovers goes into production.
  • Sila, ~$300M. Battery-materials company diversifying beyond EVs into energy storage, with customers spanning Panasonic, Mercedes and Whoop (via Equity, 2026-07-24).
  • Augustus, $180M Series B at a $1B valuation. Founder Ferdinand Dabitz is building an "AI bank," a modern correspondent bank that automates dollar-clearing and payments (the grunt work Citibank does with "tens of thousands of analysts in the wire room"). It is a conditionally approved U.S. national bank, and Dabitz says only the "eighth company since 2010" to get that nod, and it already clears billions of euros.
  • Socket, Series C at a $1B valuation (amount not disclosed). Software supply-chain security founded by open-source developer Feross Aboukhadijeh in 2021; it detects over a thousand attacks a week and protects 20,000+ organizations from OpenAI to Anthropic to Vercel.
  • Colossal Biosciences, raising new venture money at a $30B valuation. The "de-extinction" company (dodos, dire wolves) is already generating revenue, just won a government contract to save the Tasmanian devil from a contagious disease, and named its next revival target: the bluebuck antelope. Per The Best One Yet (2026-07-22).
  • Poolhouse, $55M at a $100M valuation. "Topgolf for pool," meaning AI-enabled billiards tables with tracking cameras, projected aiming lines and smart cue sticks. From the twin brothers who built Topgolf (sold to Callaway (MODG) for $2B) and then Puttshack. More on this delightful thing below.

On the AI names, 20VC spent a chunk of the July 23 episode marveling at the math of this exact moment. Harry Stebbings framed the investor's dilemma perfectly: a hot AI Series A now prices at "$300 million on $2 million in revenue," while a scaled winner like Fireworks is "less than 10x revenue" because it is about to double. As his co-host put it, "you don't really want to be a picker. You want to be a pricer." That is the whole game right now in one line.

Founder Story of the Week

Lin Qiao, founder and CEO of Fireworks, on 20VC (2026-07-20), with the $1.5 billion round landing three days later.

If you only know Fireworks from this week's raise, you are missing the better story. Fireworks sells inference, the plumbing that actually runs an AI model when you use it, as opposed to the expensive work of training the model in the first place. It sits in the middle of the stack: chipmakers like NVIDIA below, model builders above, and Fireworks in between, hosting open-source models for enterprises that want to run AI on their own data. Boring-sounding. Enormously valuable. Over $1 billion in ARR in roughly four years, and about 40 trillion tokens processed a day.

Lin's path there is the memorable part. She is a first-generation immigrant who came to the U.S. in 2000, did a PhD in distributed systems and databases, then spent years at a research lab and LinkedIn building data systems. By 2015 she thought she was ready to start a company, with the tech, the product idea, even a list of co-founders. And then she stopped herself, because she realized the missing piece was not technical:

"It's not just about product. It's not just about tech. It's actually about people. And I decided I want to go to a place I can learn the most of people. And the best company at that time is Facebook."

Her plan was to stay one year and learn how to build teams. It did not quite go to plan:

"And secretly, I was planning to learn for one year and even go back to do my own business. I stayed there for seven years."

Those seven years at Facebook are where she ended up leading PyTorch, the open-source framework that a huge slice of the AI world is built on, which is exactly why her whole thesis is a bet on openness. And she finally started Fireworks at 48, an age when Silicon Valley has usually stopped returning your calls. Her worldview is refreshingly heretical for 2026: she does not believe the future is one all-knowing model. She thinks it is the opposite:

"I really believe the future will be, it may be scary, but I think that's true, it will be millions of specialized models, one per application, per use case."

And she is blunt about the trap most AI startups are walking into, the gap between having a product people love and having a business that can survive the bill:

"For startups, we have great companies that have product market fit. Customers want to pay them and they really value their product, but they cannot scale, because once they scale, they could scale into bankruptcy. Have you heard about scaling into bankruptcy?"

That is the single most useful sentence on any podcast this week. "Scaling into bankruptcy" is the AI-era version of the dot-com trap: every extra customer costs you money because your compute bill outruns your revenue. Fireworks' entire pitch is that it makes inference cheap enough that you do not scale into bankruptcy, and Lin thinks that math only gets better: a 10x drop in the cost of a token over the next three years, driving 100x more usage.

Now the fundraising mechanics, because that is what this newsletter is for. The clearest window into how this round came together is how the last money got in. Harry Stebbings, who led 20VC into the company, does not do understatement:

"a founder who I wrote a $10 million check for after just a 15 minute meeting... This was one of the easiest investment decisions that I've made in a 10 year investing career."

There is a lovely detail about how contrarian this bet looked at the time. Benchmark's Eric Vishria, one of Fireworks' backers, apparently has a personal rule against investing in big-tech executives-turned-founders, because, as an advisor bluntly reminded him, "how many big tech executives have [been] successful in starting a company? Very few." He broke the rule for Lin. Everyone who did looks like a genius now: this week's $1.5 billion round at a $17.5 billion post-money was led by Index Ventures, Gavin Baker, Lightspeed and 20VC. On 20VC's July 23 debrief, Harry reported that Lin said gross margins are currently in the mid-30s and rising as Fireworks eats more of the stack (it plans to move down into owning data-center capacity itself), and that she expects revenue to double to $2 billion by year-end. She is also stacking her bench with heavyweights: Fireworks just hired George Hu, the former president of Salesforce (CRM).

The bear case is right there in the bull case, straight up: inference is a lower-margin, capex-hungry business, and if Fireworks has to buy or build its own data centers to control its destiny, those mid-30s margins get tested and the capex bill balloons. There is a real commoditization risk. But at "less than 10x revenue" for something growing 5x a year with expanding margins, it is easy to see why four blue-chip firms fell over each other to write the check. This is the most interesting company of the week that is not already a household name.

Also Heard

  • A venture-backed gas station. The Best One Yet (2026-07-21) flagged Maggie's Refuel, a $2 million raise for a luxury gas station concept in Beverly Hills, think artisanal snacks and EV charging, backed by ex-Starbucks and 7-Eleven executives. When boutique fuel stops are getting seed rounds, you are allowed to wonder how late in the cycle we are.
  • "A great model isn't enough." On Equity (2026-07-22), Menlo Ventures' Matt Murphy, who led Menlo's $500M+ investment into Anthropic back in 2024, walked through why he went all-in so early. His read on Dario Amodei is the classic origin story VCs dream about: "I love telling this story... Dario is a special founder... his knowledge about AI and what needs to be built and how to build it was just exceptional." Menlo's own team now runs its deal flow through Claude CoWork. The episode's thesis, and a good gut-check for founders: in this wave, a great model is table stakes, and the business around it is the real moat.
  • Poolhouse, the "rinse-and-repeat" playbook. Back to those smart pool tables. The reason this $55M round matters is not billiards, it is the founders. Twins Steve and Dave Joleff built Topgolf, sold it to Callaway (MODG) for $2 billion, built Puttshack, and are now running the exact same formula a third time: take an old game, add tech, add food and beverage, sell the mass-market upgrade. The endgame is not a chain of fancy halls; it is licensing the smart-table hardware to cruise ships and casinos and eventually selling units for ~$1,000 each. As The Best One Yet noted, second-time founders succeed at twice the rate of first-timers, and third-time founders beat even them. When a proven operator runs it back, pay attention.
  • A $3M seed lesson for the rest of us. On Pitch The PM (2026-07-23), Fiscal AI founder Shyam Nath, who raised ~$3 million across seed and a seed extension at roughly $500K ARR, gave the most practical fundraising advice of the week: use SAFEs instead of priced rounds to keep legal costs down, and once you hit ~$100K ARR, aim for a ~$10 million SAFE cap at about 10% dilution rather than reaching for a nosebleed valuation your growth cannot support.
  • Not a startup round, but worth a nod. A couple of Canadian mining financings surfaced this week, Barrick taking a strategic ~C$21M stake in Kingfisher Metals at C$1.35/share, and First Phosphate closing ~C$17.7M, but these are resource-company placements, not venture rounds, so they stay out of the tally.