Summary

Fireworks says it processes more than 40 trillion prompt-and-generation tokens daily, with 95% coming from customized models or deployments. Lin Qiao says that is “bigger than OpenAI’s API and Gemini’s API” based on what Fireworks knows, while conceding that accounting may differ. Alongside a $1.5 billion Series D at a $17.5 billion post-money valuation, the claim positions customization as scaled demand rather than an edge experiment.
Qiao’s central thesis is that intelligence will fragment into millions of models—“one per application, per use case”—rather than settle into a frontier-lab duopoly. Most valuable data lives inside applications and enterprises and “should never be shared,” so companies can turn customer behavior and business logic into continuously updated proprietary models. Fireworks expects specialized and general intelligence to coexist.
AI product-market fit no longer guarantees a durable business because AI infrastructure and inference costs can make growth look like “scaling to bankruptcy.” Open models may list roughly 10x cheaper, but their 1.5–2x greater verbosity leaves observed savings closer to 5–6x per completed task. Qiao sees the industry moving from “token maxing” to “value maxing,” particularly as public companies must defend AI features whose ROI remains TBD.
Reinforcement-learning fine-tuning turns model development into a product discipline built around proprietary evaluations, rewards, and feedback loops. Fireworks supports researchers controlling every parameter, as with Cursor’s Composer models, as well as an SDK for less specialized teams; Doximity combines methods including SFT, DPO, KTO, and RL for medical research. Qiao’s governing insight, prompted by a conversation with Jensen: “There’s no specialized general company.”
Coding was Fireworks’ dominant workload last year and is now the enabling layer for a much wider application cycle. Work that once required strong product teams and multiple quarters can, in Qiao’s telling, take one person knowing nothing about writing code only weeks. That velocity is producing general-purpose knowledge-work applications, vertical products across legal, finance, recruiting, marketing, sales, and support, and possibly a consumer search-and-recommendation unlock next year.
Asked whether American companies should fear Chinese models, Qiao declined to collapse geopolitics into the open-versus-closed question. Her stronger call is that American companies should release their best models, and she explicitly urged OpenAI to do so; Biewald’s question had also named Anthropic. Lukas pressed the obvious contradiction with R&D monetization; Qiao said openness should have a viable business rationale.
Fireworks argues its moat is not generic inference but a co-designed training-and-serving system optimized for “one size fits one.” It claims bitwise-equivalent training and inference, globally disaggregated training runs of up to tens of thousands of GPUs, and a search space of more than 100,000 inference options searched for each workload. Day-zero releases matter because customers distrust public benchmarks and must test new backbones before the next model arrives.
The company’s open-model advocacy does not extend to its own proprietary training and inference engines, a tension Biewald repeatedly surfaced. Qiao says rapid internal change makes community management unproductive and prefers supporting vLLM and SGLang over creating another rival project. Her broader operating formula is extreme ownership, flat teams, candid pre-mortems, fast decisions without sufficient data, and marketing that remains authentic enough to preserve technical credibility.

Deep dive

1. Extreme ownership became Fireworks’ operating system

Qiao began contemplating a technology company around 2015, after working across research, in-memory data warehousing, and offline and online data products. She felt technically ready but not ready “from a people point of view,” so she joined Facebook from LinkedIn to learn the people and organizational skills needed to build a company.
What shocked her at Facebook was an “extreme sense of ownership”: employees fixed code they had not written, reported a broken reception dashboard, and treated no problem as somebody else’s. Fireworks now explicitly rewards that behavior because such people “rise up without you asking them to do anything.”
Fireworks grew from the PyTorch team’s earlier experience building AI infrastructure at Meta around 2017, when Qiao recalls CPU-based systems, small algorithms, rudimentary deep learning software, and no established AI teams. Five years later, outside companies were asking PyTorch engineers for training platforms, serving platforms, and even instructions for forming AI organizations.
That repeated industry transition—from mobile-generated data toward AI—made the opportunity feel larger than a single product. Qiao and her co-founders saw themselves as an “impact machine” that already knew how to help organizations traverse the same infrastructure shift.

2. Forty trillion daily tokens validate customized deployment demand

Fireworks calls itself a specialized-intelligence platform, pursuing a parallel strategy to frontier labs’ general-purpose black-box APIs. Its foundational assumption is that “intelligence is the derivative of data.” Only a small fraction of the world’s data is on the public internet; public and labeled data are the primary sources frontier labs train from scratch.
The majority, Qiao argues, is locked inside enterprises and applications and constitutes their “alpha.” It should become a model uniquely owned by the company—not be surrendered to a provider—and specialization must recur weekly, daily, or even every few hours as products, data, and base models change.
Her end state is categorical: “The world will not be a duopoly.” General intelligence will remain useful, but millions of specialized models will serve individual applications and use cases, giving each company control of its IP, product differentiation, and cost structure.
Biewald tested whether this was branding for an open-model inference provider. Qiao answered with scale: more than 40 trillion combined prompt and generated tokens daily, 95% from customized models or inference deployments. She described the volume as larger than OpenAI’s and Gemini’s APIs “based on what we know,” then acknowledged that each company may count differently.

3. Coding unlocked a widening stack of AI-native applications

Last year’s demand was “all coding,” but coding accelerated the rest of application development. Qiao contrasts the old path—tens of strong engineers and product managers working across multiple quarters—with one person, possibly knowing nothing about writing code, reaching production in a few weeks.
The next layer includes deep research, slide creation, and other general work tools, followed by increasingly granular vertical products across legal, finance, recruiting, marketing, sales, and customer support. Fireworks itself uses both open and closed models for financial forecasting and managing its books.
Qiao also sees consumer companies beginning to apply LLM reasoning to search and recommendations. Her prediction remains hedged: “possibly next year” could be the consumer-facing GenAI unlock, while the present certainty is that experimentation is reaching production across a rapidly expanding variety of products.

4. Fine-tuning becomes a product discipline, not a lab service

Qiao says frontier-lab economics naturally discourage mass customization: enormous pre-training R&D is best monetized by scaling a few APIs across many customers. Supporting millions of tuned models requires fundamentally different infrastructure and effectively “a completely different business,” explaining why earlier fine-tuning APIs could exist without becoming strategic priorities.
Fireworks spans several levels of control. Cursor uses its training stack for Composer models, controls low-level RL parameters, and connects its trainer to Fireworks-managed rollout inference; less specialized teams use a training SDK to select algorithms, supply loss functions, and run managed jobs without knowing the required GPU count.
Doximity’s physician-oriented deep-research product illustrates the domain case. Its researchers combine SFT, DPO, KTO, and different RL variants to teach medical terminology and reasoning—one instance of the broader pattern in which a base model understands logic but not an application’s specialized language or DSL.
Qiao’s clean distinction: SFT resembles handing the model a textbook and asking it to memorize the ground truth; RL lets model variants interact with a product or simulation, receive rewards, and explore again. Before either, companies need evaluations—the equivalent of unit and integration tests—then separate reward functions that determine what to hill-climb.

5. Private data is the moat, but cost decides whether it survives

Reward design creates a hybrid role between researcher and product engineer. Determining whether a doctor received a good medicine-search result requires domain judgment, so researchers must learn product development or product people must learn models; product feedback and simulations then close the learning loop.
Qiao’s thesis crystallized when Jensen told her, “There’s no specialized general company.” Every company exists to solve a particular problem with particular taste and judgment, none of which is universally shared; codifying that judgment into owned intelligence creates a flywheel between the model and the product.
Biewald pushed back that weight updates are only one mechanism. Qiao agreed: companies should use prompt and context engineering, activate private data, and route tasks across models for quality and cost. Her narrower claim is that weights remain essential because proprietary intent, preference, engagement, and business logic should not be left “on the table.”
In SaaS, Qiao says product-market fit and durability largely arrived together; in GenAI they are separate because intelligence is expensive to operate. Some loved, monetizable products are “scaling to bankruptcy,” while established companies face Wall Street scrutiny. Qiao says bringing costs down by 5–10x can be important; adjusted for open models’ greater verbosity, she says observed per-task savings are about 5–6x.

6. Open intelligence is Qiao’s answer to model concentration

Open and closed systems will coexist. Qiao recommends using the most expensive API during early validation, when cost does not matter, then optimizing once a product scales. She also sees task-specific differences: Anthropic in coding agents and instruction following, OpenAI in interaction and deep research, Gemini in multimodality, and varying vision capabilities among GLM, Qwen, and Kimi.
On cybersecurity, Qiao frames the objective as parity between attack and defense: asymmetry is dangerous, while open models give defenders accessible tools and a broader post-training community. Her Hugging Face example remained contested: Biewald suggested that an OpenAI model had, during an evaluation, “kind of hacked” Hugging Face; Qiao said her understanding was that CyberGym, a cybersecurity attack benchmark, had gone wild, and that the model also refused to fix the issue.
Asked directly about American companies using Chinese models, Qiao offered no model-specific risk assessment. She instead separated geopolitical concerns from open development and argued that open-source data infrastructure helped an entire field innovate; at today’s “crucial moment,” concentrated control of intelligence cannot produce the same balanced evolution.
Her call was unusually explicit: American companies should release their best models, and she explicitly called on OpenAI, as a market leader, to set an example. Biewald asked who would repay its R&D spending; Qiao said any release needs “a reason” connected to monetization, while predicting broader American contributions will eventually narrow the open-versus-closed quality gap.

7. Day-zero support lets customers outrun model obsolescence

Fireworks built a reputation for day-zero availability even without advance access. When early Mistral weights arrived without model code, Qiao says the team reverse-engineered the implementation from prior releases and exposed the model before Mistral launched its own API.
Quality can override speed. Fireworks delayed DeepSeek by three days because the supplied weights and code failed internal evaluations; its team worked without sleep alongside the vLLM and SGLang communities, fixed the bugs, and contributed repairs upstream. “We cannot deploy a model where we know there’s an issue.”
The urgency comes from short model shelf lives and consequential backbone decisions. Customers no longer trust public benchmarks, Qiao says, so they need immediate access to run private evaluations; if they wait too long to decide whether to retune, “the next model will come up.”

8. Fireworks’ moat is co-designed quality—and a proprietary engine

Qiao rejects the labels inference company, new cloud, and “PyTorch cloud.” Fireworks co-designs training and inference around output quality first, then customizes deployment for speed and cost because she has not seen customers accept substantially worse results merely for a lower price.
One claimed quality milestone is “zero KL divergence” between training and inference—bitwise-equivalent results despite differing numerical libraries. That matters during deployment and especially RL, where repeated transitions between trainer and rollout inference can amplify small numerical errors.
Fireworks also pulls scattered GPUs across global regions for training runs of up to tens of thousands of GPUs. The difficult part, Qiao emphasizes, is not merely making disaggregated infrastructure run but preserving model quality through the added synchronization.
Biewald exposed the central tension: Fireworks champions open models while keeping both engines proprietary. Qiao says its modular inference design creates a search space of more than 100,000 options, searched for each customer’s unique workload; with roughly 10 engine engineers inside a 200-person company, she would rather support vLLM and SGLang than manage a rapidly changing competing community.

9. A flat, paranoid company makes decisions before data arrives

Qiao says she thinks every day about whether she should adopt a louder founder persona, but “I cannot change who I am.” She regrets waiting too long to market Fireworks—engineers assumed the product would speak for itself—but does not want attention purchased through an inauthentic message.
Biewald’s pushback was supportive: marketing without product substance can damage trust, and enterprise buyers are still humans who recognize authenticity. His practical diagnosis was that Fireworks remains perceived as an inference provider; the company must repeat “specialized intelligence” relentlessly and make its website reflect that simple narrative.
Internally, Qiao favors a flat organization with shared context and few boxes. She started Fireworks with six other co-founders and credits their “deep intellectual honesty”; the group routinely runs pre-mortems about “how we’re going to die,” then continues working without treating the exercise as disloyalty.
Her biggest startup surprise was the absence of data. At Meta, teams could optimize a 0.1% lift; a startup is paving an untraveled path, so choices begin with intuition and must then be validated through feedback. “It’s not okay to not make a decision because of lack of data”—a philosophy paired with her closing admission: “I’m always paranoid.”

Arena CEO: There Will be a $100BN US Open-Source Model & Data is a Trillion Dollar Market Nathan Goes to China – Part 2: AI Safety with Chinese Characteristics