Fireworks AI co-founder: Open models are just three to six months behind frontier models
When Fireworks AI co-founder Chenyu Zhao took the stage at the Slush conference in Helsinki, he argued that open-source models are good enough that most enterprises can skip the expensive frontier models from Anthropic and OpenAI. The same day, Anthropic completed history's largest IPO, and Fireworks was still working through two security incidents in the fall that exposed data at customers Uber and Notion. That makes the company a concentrated picture of the two opposing pressures in enterprise AI right now: cost pressure pulling toward open models, and operational risk pulling toward the established labs.
What Zhao argued
Fireworks AI runs AI models for other companies. The company specializes in open models — models with open weights that anyone can download — and helps customers fine-tune them on their own data, so they run faster and cheaper than the models from Anthropic, OpenAI and other heavyweights, while also becoming harder for competitors to copy. Its customers have included Uber, Shopify and Notion.
Zhao's core number is the time gap between open and frontier models: "Today it is at most six months, probably closer to three," he said on stage, according to TechCrunch via MSN. This is a company estimate, not an independently verified result — Fireworks has a commercial interest in the gap appearing small.
Zhao also drew a concrete threshold: Companies spending between $10,000 and $100,000 a month on tokens should consider open source. Again, this is the company's own recommendation, not an analyst benchmark.
The two-model thesis: cost per task, not per token
Zhao predicted the market will consist of two types of models in parallel. According to TechCrunch, he argued that we are going to live in a two-model world, with frontier models and open models side by side: You want the best intelligence for advanced planning, but you don't need a frontier model for simple arithmetic. That is why there is so much interest in routing, and why people are now thinking about cost per task, not just cost per token.
The logic is simple: If most of an enterprise's AI tasks are routine work, and only a small share requires top-tier intelligence, routing between the model types can cut the token bill significantly. This is also where Fireworks' business model fits in — the company delivers the infrastructure that makes fine-tuned open models fast enough to compete in those routine-based tasks.
The counterweight: two security incidents in the fall
The same company has, according to its own information as reported by TechCrunch, had two security incidents this fall:
- In one incident, hackers reached API keys and tokens stored as secrets for a limited number of accounts.
- In the other, attackers broke into the company's cloud service provider, made their way into a development system, and obtained customer data, including billing records and some datasets.
The consequences for customers are documented. Uber says that names, addresses, birth dates and driver's license numbers of around 150,000 drivers were stolen via a vendor that TechCrunch identified as Fireworks. Notion says that AI chat transcripts from fewer than 1 percent of its workspaces were exposed, and it has disconnected from Fireworks.
The incidents show the practical side of Zhao's own pitch: When companies gather sensitive data and fine-tuned models with a custom-infrastructure vendor, a new exposure surface emerges that does not exist when you simply send tokens to an established lab. It is worth noting that the frontier-model camp has its own counterargument in the same direction — Microsoft CEO Satya Nadella raised concerns in July that the frontier labs absorb too much of customers' know-how, even if only through the prompts used to get answers. Both paths thus carry documented risks; Fireworks' is currently more visible.
The IPO party that sets the frame
Zhao's argument came the same day Anthropic completed an IPO that, according to The Motley Fool, raised $81.3 billion — $6.3 billion more than SpaceX's $75 billion listing, making it history's largest measured in capital raised. The shares were priced at $83 and closed the first day at $97.17, roughly 17 percent above the offer price, according to BBC.
The point is not that Anthropic enthusiasm disproves Zhao's cost thesis. But it shows how much capital and faith now sit on the frontier-model side of the debate — and how large the difference is between claiming the gap is three months and the market believing it.
A market moving fast
Interest in the custom-model side is also documented in money flows. On Wednesday, Palo Alto Networks announced it is buying Applied Compute for $4.5 billion, the entire amount in stock. Applied Compute, which like Fireworks helps companies train and run their own AI models on their own data, was founded just a year ago by former OpenAI researchers. Rivals Together AI and Baseten have announced large funding rounds over the past few months.
Fireworks itself was given a valuation of $17.5 billion by its investors this summer, and is, according to TechCrunch, in "early talks" about a new round. Zhao declined to comment — the round is thus reported, not confirmed.
What remains open
Several things should be read with caution. First, all of Zhao's numbers — the six-month gap, the token threshold, the two-model thesis — rest on a stage appearance and the company's own framing, not independent benchmarks. Second, the security incidents are known through TechCrunch's reporting of Fireworks' own disclosures; the company's primary incident reports and the exact scope of the Uber and Notion exposure are not directly documented here. Third, the funding round is uncertain until confirmed.
The question that remains is nonetheless concrete: If Zhao is right about a three-month gap and routing by cost per task, enterprises win with open infrastructure. If the fall incidents are the symptom, they pay for it. Most likely, both are true at the same time — and the choice between frontier models and open models becomes a choice between two different risk profiles, not a question of whether one of them is safe.

