Key Takeaways
- • Fireworks AI co-founder Chenyu Zhao argued on the day Anthropic went public in the largest IPO on record that companies can avoid frontier-model prices by adopting open-weight models tuned on their own data.
- • Zhao said open models are at most six months behind frontier models, probably closer to three, and that companies should consider switching once they reach 10,000 to 100,000 users and spend $10,000 to $100,000 a month on tokens.
- • Fireworks disclosed two security incidents this fall; Uber says the names, addresses, birth dates and driver's license numbers of about 150,000 drivers were stolen through a vendor TechCrunch identified as Fireworks, and Notion disconnected from the company after AI chat transcripts from fewer than 1% of its workspaces were exposed.
- • Fireworks, valued at $17.5 billion in summer, is reportedly in early talks for another funding round, though Zhao declined to comment.
- • Palo Alto Networks agreed to buy Applied Compute, a rival custom-model platform founded a year ago, for $4.5 billion in stock.
Summary
Zhao argued that frontier models are expensive and usually unnecessary unless a task requires the latest general intelligence. He said open models have narrowed the gap considerably, from roughly a year or a year and a half a few years ago to at most six months today, and he expects a "two-model world" in which frontier and open models run in parallel, with routing and cost per task becoming the key metrics. He also noted that enterprises often start with frontier models because they are a low-risk brand choice before reaching scale, and that concerns about labs absorbing customers' know-how, which Microsoft CEO Satya Nadella raised in July, are growing.
Fireworks has also faced scrutiny over two security incidents disclosed this fall. In one, hackers reached API keys and tokens stored as secrets for a limited number of accounts. In the other, attackers breached the company's cloud provider, entered a development system and obtained customer data, including billing records and some datasets. Zhao said the first incident was driven largely by "LLM jacking," in which attackers hunt for API keys to resell tokens on illicit marketplaces, and that the company's zero data retention promise for inference traffic has no indication of compromise. He acknowledged that Notion's batch prediction workflow stored data at rest, which is why the company is moving those services to an online flow. Uber has said about 150,000 drivers' personal and license data were stolen through a vendor TechCrunch identified as Fireworks.
The company is valued at $17.5 billion following its summer funding round and is reportedly in early talks for another round. Competitors are also attracting capital: Together AI and Baseten have announced large raises, and Palo Alto Networks said it will acquire Applied Compute, founded a year ago by former OpenAI researchers, for $4.5 billion in stock. Zhao also said Nvidia, an investor since 2024, remains committed to open source following the departure of CEO Jensen Huang last week.
Go deeper than the headline
Research every firm in this story, track the ones that matter and get the daily briefing — free for 7 days.