AI & ComputeLOT 261120 · 5 min read
Fireworks AI says open models trail frontier models by months
Fireworks AI, valued at $17.5 billion, says open-weight models trail frontier models by months and calls cost per task the deciding metric, after two disclosed breaches exposed Uber driver data.
Key takeaways
- Fireworks AI disclosed two security incidents this fall; Uber says personal data on about 150,000 drivers was stolen through the vendor TechCrunch identified as Fireworks.
- Fireworks investors assigned the company a $17.5 billion valuation in the summer, and it is reportedly in early talks for another round.
- Palo Alto Networks said on Wednesday it will buy Applied Compute, a year-old startup founded by former OpenAI researchers, for $4.5 billion in stock.
- Zhao puts the open-model gap at six months behind frontier models at most, and closer to three months in practice.
- Zhao cites 10,000 to 100,000 users and $10,000 to $100,000 a month in token spending as the point where open source becomes worth considering.

Fireworks AI, identified by TechCrunch as the vendor through which hackers stole personal data on roughly 150,000 Uber drivers, disclosed two security incidents this fall, and the open-weight model platform still carries a $17.5 billion investor valuation. What Fireworks sells runs against the grain of the AI business: models anyone can download, tuned on customer data so they run faster and cheaper at inference.
Fireworks runs those open-weight models for other businesses and helps customers tune them on their own data, which makes them harder for rivals to copy. Uber, Shopify and Notion have used the platform. "I describe Fireworks as a specialized intelligence platform," co-founder Chenyu Zhao said. "We use open models, which are often smaller, cheaper and faster, and let companies add their own data to specialize them."
Zhao spent a decade at Google, including a 50-person engineering team on Google Cloud Vertex AI, before founding Fireworks with Lin Qiao and Dima Dzhulgakov. He said the decision took five minutes, once he met the pair and found them both genuinely talented and decent company. The pitch he brings to customers is ownership of the model itself, "rather than being beholden to giant AI conglomerates."
How far behind are open models now?
Zhao put the distance between open and frontier models at six months at most, and nearer three in practice. A few years ago he would have said a year and a half, until DeepSeek arrived and a run of strong open releases followed. The gap no longer spans generations, which changes the buying decision for any company paying frontier prices for routine work.
Switching still lags because frontier models are household brands, and before a product finds its market the best model remains the low-risk choice. The trigger, Zhao said, arrives past 10,000 to 100,000 users and $10,000 to $100,000 a month in token spending. Beyond that line, he said, it is time to think about open source.
Zhao expects a two-model market rather than a winner-take-all one: frontier systems for high-level planning, open systems for everything else, including basic arithmetic. That has made routing a product category, and cost per task a purchasing metric rather than cost per token. In July, Microsoft CEO Satya Nadella warned that companies on proprietary models pay twice, once in money and again in know-how. Zhao repeated the concern from the Slush stage in Helsinki, where open-source tooling dominated the agenda.
What did the two incidents expose?
The incidents covered different ground. In one, hackers reached API keys and tokens stored as secrets for a limited number of accounts. In the other, attackers breached the cloud provider Fireworks uses, reached a development system and took customer data, including billing records and some datasets.
Notion reported that hackers exposed AI chat transcripts from fewer than 1% of its workspaces, and Notion disconnected from Fireworks. Cursor has also gone public about its own exposure. Uber said the stolen records covered names, addresses, birth dates and driver's license numbers for about 150,000 drivers, through a vendor that TechCrunch identified as Fireworks.
Zhao called the pattern an industry condition. "Every company is probably seeing an unprecedented frequency of attacks," he said, adding that "a lot of the motivation was LLM jacking: hunting through providers for API keys and secrets to sell tokens on illicit marketplaces." Fireworks promises zero data retention for inference traffic, and Zhao said no sign of that promise being broken has surfaced. Notion had used batch prediction, where uploaded datasets sit at rest, which Zhao conceded is not a zero-retention product by nature. The company is moving those services to an online flow that keeps data only as long as it needs it.
Capital is chasing the category. Fireworks investors assigned a $17.5 billion valuation in the summer, and the company is reportedly in early talks for another round, a matter Zhao declined to comment on. Together AI and Baseten have announced large rounds in recent months.
On Wednesday, Palo Alto Networks said it will buy Applied Compute, a startup founded a year ago by former OpenAI researchers that helps companies train and run custom models on their own data. The $4.5 billion price, entirely in stock, is roughly a quarter of Fireworks' valuation for a company one year old.
Anthropic went public the same day, in the largest IPO on record, with investor enthusiasm for the AI lab running high. Zhao declined to pick a winner between two futures: Anthropic cementing its place in an oligopoly, or Anthropic losing traction. On the record, he said he would probably call his broker that night and buy a few shares, calling it a long-term bet on AI usage reaching a scale that demands immense capital.
For chip and infrastructure suppliers, the commercial question is which models customers host and tune themselves, and how much inference that moves. Nvidia has backed Fireworks since 2024, and Zhao, asked about that relationship in the context of Jensen Huang's departure from the chipmaker last week, said he expects Nvidia to keep investing in Nemotron and to remain a strong supporter of open source.
Zhao expects frontier and open models to run in parallel for years, with routing and cost per task deciding the bill. The price fight now runs between the labs and the platforms, and open-weight serving keeps closing the distance.
techcrunch.com (Source)
More from Nathan Brooks
Show full bio
Senior reporter covering people, launches and day-to-day practice in Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.
1,142 articles