Open-weight AI models are now close enough to the frontier that many companies can skip the priciest options, according to Fireworks AI co-founder Chenyu Zhao. Fireworks runs open-weight AI models for businesses and is betting that enterprises will stop paying frontier prices for jobs that don’t need them.
He spoke at the Slush startup conference in Helsinki, the day Anthropic listed in the biggest IPO ever recorded.
What Fireworks sells
Fireworks hosts open-weight AI models, which anyone can download, and helps customers tune them on their own data. The result runs faster and cheaper than models from Anthropic, OpenAI and other large labs, and is harder for rivals to copy. Past customers include Uber, Shopify and Notion.
Zhao calls Fireworks a specialized intelligence platform. In his framing, OpenAI and Anthropic sell general intelligence, while Fireworks starts from smaller open models and lets a company mold them around its own data, so the company owns the result. Using the same Claude or GPT models as every competitor, he argues, gives a business nothing that sets it apart.
How far behind open models are
Zhao puts the gap at six months at most, probably closer to three. He once would have estimated a year to eighteen months. DeepSeek and the strong open releases that followed shrank that distance.
Why do firms hesitate? Zhao says a famous frontier model is the low-risk choice while a product is still looking for customers. His threshold for reconsidering comes when usage reaches 10,000 to 100,000 users and monthly token spending reaches $10,000 to $100,000.
A two-model world
Zhao expects both kinds of model to coexist. Top-tier intelligence makes sense for high-level planning, but simple calculations don’t need it. He says that’s driving interest in routing each request to a suitable model, and in tracking what a whole task costs rather than the price per token.
Data plays a part too. In July, Microsoft CEO Satya Nadella cautioned that firms relying on proprietary models hand over money and also their accumulated know-how.
Two security incidents this fall
The pitch comes with a problem. Fireworks has disclosed two security incidents this fall. One exposed API keys and tokens held as secrets for a limited number of accounts. In the other, intruders got into its cloud provider, then a development system, and took customer data such as billing records and some datasets.
Notion says chat transcripts from AI features in fewer than 1% of its workspaces were exposed, and it has cut its connection to Fireworks. Uber says about 150,000 drivers had names, addresses, birth dates and license numbers stolen through a vendor.
Zhao says much of the motive was “LLM jacking,” where attackers hunt for API keys to resell tokens on illicit marketplaces. He says Fireworks promises zero data retention for inference traffic and sees no sign that was compromised. Notion’s data came from batch prediction, where customers upload datasets that sit in storage, so that service isn’t zero retention by design. Fireworks is moving it to an online flow that keeps data only as long as needed. He wouldn’t name other affected customers, though Cursor has said it was among them.
Money keeps arriving
Investors valued Fireworks at $17.5 billion in the summer, and it’s reportedly in early talks for another round. Zhao declined to comment on that.
The category is crowded with cash. Together AI and Baseten have each announced large rounds in recent months. Palo Alto Networks said Wednesday it will acquire Applied Compute for $4.5 billion in stock. The startup, founded a year ago by ex-OpenAI researchers, also helps businesses train and run custom models on their own data.
What to watch next
Zhao said Nvidia has backed Fireworks since 2024 and that he expects it to keep investing in its Nemotron open models, even after Jensen Huang’s departure from the company last week. He called Huang’s exit a blow to the industry.
The test for Fireworks is whether customers who left after the breaches come back, and whether its pitch on cost holds as frontier prices move. Businesses weighing open-weight AI models should compare cost per task and ask hard questions about how a host stores their data.
Source: TechCrunch