Fireworks AI co-founder Chenyu Zhao made the case for open-weight models at Helsinki's Slush conference: they cost less than Anthropic or OpenAI's services and let enterprises customize AI on their own data without exposing proprietary information to larger rivals.
The pitch has clear economic logic. Frontier models are expensive and overkill for most tasks. Open-weight alternatives—freely downloadable, malleable with customer data—promise to become harder for competitors to replicate once trained. Fireworks' customers have included Uber, Shopify and Notion.
But the company's $17.5 billion summer valuation now carries a large asterisk.
Two security incidents this fall exposed the core tension in Fireworks' model. In the first, hackers stole API keys and tokens from a subset of customer accounts. In the second, attackers breached Fireworks' cloud provider, accessed a development system and exfiltrated customer data including billing records and datasets.
Notion disclosed that AI chat transcripts from less than 1 percent of its workspaces were exposed and subsequently disconnected from Fireworks. Uber confirmed that hackers accessed the names, addresses, birth dates and driver's license numbers of approximately 150,000 drivers through the breach.
Those losses test whether the open-source cost advantage can overcome customer concerns about data security—a worry that cuts both ways. Microsoft CEO Satya Nadella raised in July a separate risk: that frontier labs absorb customer intelligence through prompts alone.
Fireworks is reportedly in early discussions for another funding round, though Zhao declined to comment. Capital continues flowing into the category—Together AI and Baseten announced major raises recently, and Palo Alto Networks paid $4.5 billion in stock Wednesday for Applied Compute, a year-old firm founded by former OpenAI researchers that also helps companies train custom models on proprietary data.
