Technology

Fireworks AI Says Most Companies Can Skip Frontier Models, Then Its Security Record Gets Tested

Martin HollowayPublished 2m ago4 min readBased on 2 sources
Reading level
Fireworks AI Says Most Companies Can Skip Frontier Models, Then Its Security Record Gets Tested
source:fireworks.ai

Fireworks AI co-founder Chenyu Zhao told TechCrunch on stage at the Slush startup conference in Helsinki that most companies do not need frontier models. He said frontier AI models are expensive and, in most cases, unnecessary unless a company needs the very latest generalized intelligence. He made the remark on November 19, 2026, the day Anthropic went public in what TechCrunch described as the largest IPO on record.

Zhao framed Fireworks as a specialized intelligence platform, in contrast to OpenAI and Anthropic, which he described as building general intelligence. The pitch will be familiar to anyone who has managed inference budgets: run a smaller model tuned to a specific task rather than paying for a general-purpose system on every call.

Fireworks runs AI models for other businesses. Its focus is open-weight models, whose weights are publicly available, which customers fine-tune on their own data so they can run faster and cheaper than models from Anthropic and OpenAI. Its blog presents the offering with the line "Build your own frontier. Start running open models in minutes." Customers named in coverage include Uber, Shopify and Notion. Its investors assigned the company a $17.5 billion valuation in summer 2026.

The company has also shipped product through the year. On August 31, 2026, Fireworks announced that its Training API was generally available, meaning it was open to all customers. On November 3, 2026, it announced new products at its Forge event, aimed at helping companies build specialized intelligence. Zhao spent a decade at Google, including managing a 50-person engineering team on Google Cloud Vertex AI.

The security record is the harder part of the story. Fireworks disclosed two security incidents in fall 2026. In the first, hackers reached API keys and tokens, the credentials that let software call a service, which were stored as secrets for a limited number of accounts. In the second, attackers breached the company's cloud provider, entered a development system, and took customer data, including billing records and some datasets.

The downstream effects reached customers. Notion said AI chat transcripts from fewer than 1% of its workspaces were exposed in the Fireworks incident, and that it disconnected from Fireworks. Uber said hackers stole names, addresses, birth dates and driver's license numbers of about 150,000 drivers through a vendor that TechCrunch identified as Fireworks AI. The Fireworks blog lists SOC 2 Type 2, ISO 27001, ISO 27701 and ISO 42001 certifications, which are independent audits of security controls.

In my view, the combination of certifications and incidents is the detail enterprise buyers will weigh most heavily. A compliance attestation describes controls at a point in time, while the second incident points to a failure in a dependency chain, namely the cloud provider and a development environment holding customer data. For a platform whose value proposition is custom models trained on proprietary data, the question of where that data sits, and who can reach it, is central to the sale.

The broader context here is a shift in procurement logic. If many production workloads can run on tuned open-weight models at lower cost, the frontier labs will compete for the smaller set of tasks that require general capability, while infrastructure providers compete on cost, latency and the operational discipline around training data. That is a more crowded market, and it rewards security practices customers can verify. Zhao's argument that most enterprises do not need the most capable model will be tested over the next several quarters. The Fireworks incidents are a reminder that the plumbing around a model matters as much as the model itself.