Fireworks AI CEO Lin Qiao: A 10x Token-Cost Drop Drives 100x Usage

Cover image

The Fireworks AI CEO processes 40 trillion tokens a day and thinks the capex-bubble crowd has the arithmetic backwards. This is her on Harry Stebbings' 20VC talking about how her real push is about who gets to own intelligence.

The short version, if you don't have the hour:

Fireworks processes more than 40 trillion tokens a day, most of them from customized models rather than off-the-shelf ones, and Qiao expects that volume to grow 20 to 100x by the end of next year on what she calls the early stage of an S-curve.
She sees token costs falling roughly 10x over three years, and that drop pulling usage up about 100x. Costs haven't dropped yet because of a supply-chain shortage in energy, chips, and manufacturing rather than any ceiling on the technology.
Gross margins of 30 to 40 percent, against the 80 percent a SaaS business earns, are a hypergrowth choice rather than the new normal. Fireworks won't move into the application layer, and moving down into data centers is a question of timing.
Revenue is near $800 million dollars of ARR and she expects it to at least double by year end, with 200 people.
Frontier labs like OpenAI and Anthropic are building a power line to distribute intelligence, foundational and valuable, but in her read they will not replace the specialized intelligence every company builds on top. She does not want a world where one company owns intelligence.
Open and closed models have both crossed a quality threshold. Open weights give control, are far easier to tune, and a small amount of a company's unique data can hill-climb an open model past a general-purpose one on that company's own task.
Last year was the year of coding, and this year is the year of co-work, a far more diversified set of jobs across legal, finance, support, recruiting, sales, and healthcare.
Hardware used to depreciate over six years. Now one vendor ships three chip generations a year and models turn over weekly, so the pace of model development outruns hardware depreciation and rewrites the build-versus-buy math.

Fireworks AI sits in the middle of the AI stack, below the model labs and above NVIDIA and the other chip makers, and that position usually invites the accusation of being a commodity pass-through. Lin Qiao, the co-founder and CEO, spent the hour with Harry Stebbings arguing the opposite. Her case starts with where intelligence actually comes from, and it ends with a warning about concentration that most infrastructure founders wouldn't say out loud.

The layer that turns locked data into intelligence

Qiao treats intelligence as a derivative of data, and then points out where the data lives. Frontier models train on the public internet plus labeled data, and by her account that is the small corner of what exists. If intelligence is a derivative of data, most of the world's data never gets used to train a general intelligence model at all. The public internet is small next to the world's total, and the majority of that total is private data locked inside enterprises, which will never share it because it's proprietary IP. Her conclusion follows from there. The bulk of the world's data never gets activated into intelligence at all, and activating it is the business.

> "Majority of data is not being activated to derive any intelligence, and that's where we believe is to activate that data. We believe the future of the frontier of the intelligence is actually private intelligence, or specialized intelligence." > > Lin Qiao > > ▶ Watch · 00:05:22

Stebbings pushed the obvious objection, which is that Anthropic's enterprise business, Claude and its adjacencies, is aiming at exactly this private-data prize. Qiao's answer separated the two. A company that fully believes in AGI believes one model will eventually solve every problem in the best way, which means specialization becomes unnecessary by definition. Fireworks is a wager in the other direction.

Why she doesn't believe in one model to rule them all

Qiao's objection to the single-model future is less technical than civilizational:

> "The definition of AGI is there's this one model that can solve all the problem in the best way. To me, that means you do not need to specialize. If our future world is going to be ruled by one standard, a taste dictated by one company, we turn ourselves into an army of robots. And that's very depressing to me." > > Lin Qiao > > ▶ Watch · 00:06:02

She grounded the point in a conversation with Jensen Huang recorded after his GTC keynote, and it shapes her worldview. His version, as she relayed it, was that there is no such thing as a specialized general company, because every company is built on a special belief about how to do things that gives it a reason to exist, and that belief is not learnable or shared by another company sitting outside. Later in the hour she named the fear directly, and it is the closest thing the interview has to a thesis statement.

> "What I don't want to see is there's only one company owns intelligence. There's this general common intelligence that benefits everyone, and then there's a specialized intelligence that actually helps us advance in history. I don't want that to die because there's only one company can do that." > > Lin Qiao > > ▶ Watch · 00:29:40

I read this as a product strategy dressed as a values statement, and the two reinforce each other. If specialization is where ongoing advantage lives, then Fireworks selling the tools to specialize is selling picks in the more defensible half of the market.

Power lines are infrastructure the economy still builds on

If AGI isn't inevitable, why do the most credible founders in the field, Dario Amodei, Sam Altman, Larry Page, Sergey Brin, talk about it as though it is? Qiao doesn't dismiss them. She recasts what they're building.

> "I think what they built is fantastic, because they are basically building a power line to distribute a really great source of intelligence that everyone else can build on top of. But the question is, is this power line going to replace everything we do? I don't think so." > > Lin Qiao > > ▶ Watch · 00:10:05

Stebbings sharpened it into the investor's version of the question, whether power lines are good businesses in a world where open source is closing the gap. The answer took the rest of the hour, and it began with why Fireworks built on open weights in the first place.

Open and closed both crossed the quality line

At founding, the Fireworks team debated whether to train their own models or build on open ones, and open models were then in their infancy. Their PyTorch background settled it, and the reasoning has aged well. Openness gives control to the user, since once a model is released the weights are fully yours, and that call paid off as both open and closed models crossed a quality threshold over the past two years. The second half of the argument changes unit economics. An open model is not only controllable, it is steerable with very little data.

> "Open model cross a threshold is so much easier to tune. With a small amount of unique data a particular company has, we can hill climb towards your eval, and often time the end result of hill climbing is you solve your unique problem with your data better than a general purpose model." > > Lin Qiao > > ▶ Watch · 00:13:56

Fireworks eats its own cooking here. Qiao said the company runs open models internally for recruiting, candidate sourcing, some finance processes, and coding and debugging, in part because it is cost-conscious itself.

Scaling to bankruptcy

The overvaluation debate, according to Qiao, is really a question about whether product-market fit still equals a durable business. In the SaaS era it did. Compute was a commodity and the hard part was finding demand. Now the two have come apart.

> "During SaaS time, product market fit and the durable business almost are equivalent. And now they are two separate concepts. Customers want to pay them, but they cannot scale, because once they scale, they could scale into bankruptcy. Have you heard about scaling to bankruptcy?" > > Lin Qiao > > ▶ Watch · 00:15:10

The trap is worse for incumbents than for startups, because incumbents already have the traffic they would have to serve. Once an incumbent rolls out AI features it reaches its entire customer base at once, and the CFO's cost forecast leaves no way to justify the bill. The alternative Qiao points to is taking control of an open weights model and rolling out your own instead.

The read-across is that the frontier labs are not overvalued so much as pointed at a narrower slice of demand than the headlines imply, because a large share of enterprise workflows can be served more cheaply by a tuned open model. That's the wedge Fireworks sells into.

Put your own guardrails on every model

The most current risk Stebbings raised was national security, since the top open models on OpenRouter are now Chinese and, in his words, incredible quality. Qiao didn't brush off the concern, and she didn't treat it as unique to Chinese models either.

> "It doesn't matter if open or closed, you should put your own guardrail around it. A model provider will infuse their own judgment, their own taste, into the model training process. You cannot guarantee it matches yours." > > Lin Qiao > > ▶ Watch · 00:20:06

From there she drew the forecast that gives the interview its shape, and she conceded it might sound alarming before making it.

> "I really believe the future will not be a few small number of AGI models dominate the world. I really believe the future will be, it may be scary, but I think that's true, it will be millions of specialized model, one per use case." > > Lin Qiao > > ▶ Watch · 00:20:59

Asked whether China itself might restrict access to its open models given how fast they're improving, she granted a big short-term impact and then leaned on the structure of open ecosystems. An open ecosystem is by definition not one provider, and on talent density and resources she believes the US can and should build that open system itself.

Owning intelligence in the workflow layer

Stebbings pressed the debate through two legal-AI companies he knows well, one that committed to building its own model and one that didn't. A year ago the no-own-model camp looked right, because frontier capability was rising so fast. Now it looks wrong. Qiao's answer started with how much the ground has shifted under application software. General intelligence for coding has significantly disrupted software development, collapsing what once took tens of very strong product engineers and PMs and multiple quarters or years of investment into something one person can launch into a product and scale quickly in a few weeks. When implementation collapses as a barrier, the defensibility has to move somewhere, and she located it in the harness, the layer that decides which tool the model calls and when.

> "There's a harness integrating and deciding, orchestrating which AI tool to use, which tool to call to. And this is bespoke, this is customized... And even that harness need to be co-trained with the model powering it. So they're... driving that business to excellence by owning their own intelligence of how to do that in the workflow layer." > > Lin Qiao, on the legal-AI companies > > ▶ Watch · 00:25:57

She was careful about the counterpoint. Legal is conservative and intolerant of errors, and enterprise sales cycles run multi-year, so a raw model is nowhere near a deployment. The value both companies own is proprietary knowledge of how legal work is actually done, and that is what a co-trained harness encodes.

Model IQ moves in steps, specialization compounds

The fear that model superiority is permanently transient, one lab on top today and another tomorrow, gets a two-layer answer. Base intelligence advances in discrete jumps. Specialization on top of it moves faster and more continuously.

> "There's base general IQ advancement, and those will take step functions... I see those as every year or every three quarters there's a major leap. But build on top of the best base models, the specialization start to accelerate... So I do see specialization is going to accelerate much faster than the general intelligence part." > > Lin Qiao > > ▶ Watch · 00:27:18

The strategic read is that a company whose thesis rests on base-model supremacy is tied to a variable that jumps once a year, while a company compounding specialization improves every week. Qiao extends the same logic to routing. In her view the frontier is not one model but the routing system that sends the hard judgment to an expensive closed model and the sub-tasks to cheaper open ones, and she sees room to build an automatic version that learns by itself, a self-evolving system shaped by what flows through your product.

Inside Cursor's training engine

Fireworks' work with Cursor shows the strategy in operation, and its CTO Dima was embedded at Cursor for months building reinforcement-learning infrastructure. The starting problem is that both companies are capital-conscious startups that can't afford a hyperscaler's training cluster, so they had to make distributed training numerically sound. Cursor's work is massive reinforcement-learning post-training, where a trainer tweaks the model's weights and each new model deploys to an RL rollout that interacts with a synthetic coding environment and gets a reward back, and Fireworks decouples the trainer from the rollout. The payoff is running the same job across scattered GPUs instead of one monster cluster wired together with InfiniBand, at the cost of a hard synchronization problem.

> "In large hyperscaler, they run that all together, 10,000, 100,000 chips interconnected through InfiniBand. It's extremely expensive. We design a fully distributed system, we run across five, six data center regions globally and tap into scattered GPUs. The challenge is we need to sync model weights, because the delay dictates how fresh the rewards are." > > Lin Qiao > > ▶ Watch · 00:36:10

On the concentration risk, that Cursor is one giant customer, Qiao acknowledged the industry is bimodal, with only a few apps reaching escape velocity, and said the customer base has diversified since the year when every model company crowded around Cursor.

From the year of coding to the year of co-work

The line that will travel from this interview is her calendar for AI adoption, and it points at where the next wave of demand is coming from.

> "Last year is the year of coding. All major coding companies are on us. And this year is the year of co-work. And co-work is much more diversified by itself than coding, because there's general purpose co-work to help you do all kind of research, and so many different categories of special purpose co-work: legal, finance, customer support, recruiting, sales, marketing, healthcare." > > Lin Qiao > > ▶ Watch · 00:38:37

The commercial significance is that coding concentrated spend in a handful of winners, while co-work spreads it across a much wider field of applications, which is a more forgiving demand base for a platform that sells to all of them. She also flagged an early uptick in consumer-facing companies rethinking traditional recommendation systems through generative AI, a market she watched up close at Meta.

40 trillion tokens a day, and why the bubble talk is backwards

The numbers that anchor the whole discussion arrived together. Fireworks processes more than 40 trillion tokens a day, most of them from customized models rather than off-the-shelf ones, and Qiao expects the volume to compound off a low base.

> "Anywhere ranging from 20 to 100x could be possible. We're at a very early stage of S-curve of explosion right now." > > Lin Qiao, on token volume by the end of next year > > ▶ Watch · 00:41:22

If demand really grows 20 to 100x, Stebbings put it, the idea of a capex bubble looks absurd and the industry is under-provisioning compute. Qiao agreed, and located the real constraint at the bottom of the stack rather than the top.

> "I think Jensen has a five-layered AI cake, from top down: application, model, infrastructure, chips, energy. We are bottlenecked by the lower part in terms of supply chain. Being energy, being chips, how fast we can manufacture. All this industry is not designed for massive scaling." > > Lin Qiao > > ▶ Watch · 00:41:40

The caution I would attach is that a founder whose revenue rises with token volume has every reason to see under-provisioning rather than over-provisioning. Her supply-chain point stands on its own though, since energy and fabrication capacity are physical constraints that no funding round fixes on a quarterly timeline.

Specialize, or earn the right to build

With OpenAI, Anthropic, DeepSeek, and Meta all designing their own chips, Stebbings asked whether Fireworks has to become full-stack to survive. Qiao's answer was a discipline about focus, and a rule for when vertical integration is earned.

> "To us, agility is everything, and we need to earn the rights of building anything. We want to run everywhere, all possible AI chips in the world. But when the business grows very big, you earn the rights to build for your own giant traffic, and if it saves like five times more cost, then you should go do it." > > Lin Qiao > > ▶ Watch · 00:43:22

The natural follow-up is why NVIDIA builds an open model, Nemotron, but not an inference platform that would compete with Fireworks. Her explanation makes NVIDIA's model work a defensive supply-chain move rather than a land grab. The model, in her read, is purely a supply-chain answer, since the US not having a native open model would be a problem, and once a company like Fireworks supplies the specialized intelligence platform layer, that gap closes and NVIDIA has no reason to build the platform itself.

The cost curve that reshapes everything

Stebbings brought a hard number from Marc Benioff, a new investor, that Salesforce spends about 3.8% of its developer salaries on Anthropic and Claude Code, and asked whether that share rises or falls from here. Qiao expects the unit cost to fall hard, and explained why it hasn't yet. It hasn't come down so far because of the supply-chain constraint, but in a free economy a high price invites competition, competition brings the cost down, and much more affordable infrastructure invites more usage. She resisted a single headline multiple, because not all tokens cost the same to produce or consume. A model that is 2x cheaper per token but needs 2x more tokens to solve the same task lands at the same cost. Pressed for a figure anyway, she gave the one that headlines the episode, built from three compounding levers: fewer tokens per task through precision and tuning, better economics per token through the platform, and cheaper underlying infrastructure as the supply-chain constraint eases over two to three years.

> "So overall, I can imagine 10x cost reduction in the next three years. And this 10x cost reduction will drive 100x usage." > > Lin Qiao > > ▶ Watch · 00:50:29

Her 100-to-10 ratio is the claim anyone underwriting AI demand has to price. If cost falls 10x and usage rises 100x, revenue on tokens still grows even as the price per token collapses, which is the reconciliation the bubble debate keeps missing. On why Fireworks prices above cheaper rivals like Together, Qiao argued the comparison isn't apples-to-apples, because most of her traffic is customized models tuned for quality, down to a technical guarantee she called zero KLD, bit-equivalence between the training system and the inference system so no accuracy is lost when a model moves between them.

Why 30 to 40 percent margins are a deliberate choice

Stebbings named the number that worries SaaS investors, that Fireworks likely earns 30 to 40 percent gross margins against the 80 percent a software business earns, and asked whether that is permanent. Qiao framed it as a deliberate trade during a specific phase.

> "It is a reflection of we are in a hypergrowth phase. Margin optimization is a constraint problem, and usually constraints slow down innovation. Once we know this is a system we want to scale a thousand times bigger, then we go optimize the heck out of it." > > Lin Qiao > > ▶ Watch · 00:54:15

On where the margin expansion comes from, she drew a firm boundary at the top of the stack and left the bottom open. Moving into the application layer is off the table and clear to her, while moving down the stack into building data centers stays on the table as a question of timing.

Her rule for the growth-versus-margin question, when Stebbings asked how she decides, was to optimize product differentiation continuously while refusing to over-optimize margin into slow growth. She built data centers at Meta and sees room to specialize there, including combining flops-heavy GPUs with SRAM-heavy accelerators for the two different phases of running a model. Data centers, according to her, are not commoditized at all, since construction, power, fiber, and liquid cooling are deep operational skills.

Hardware that ages faster than it depreciates

Depreciation gives the most investable observation of the hour, and it rewrites when owning compute pays off. Hardware still physically lasts six years. The economic life is now much shorter, because the models that run best on it turn over in weeks.

> "In the past, hardware depreciation is six years, solid six years. And now, within a year, from one vendor alone, we have three SKUs. Every week we are launching a new model, and the model is peaking its value before the next model comes out. Which model runs on the two-years-old hardware? It will be a two-year-old model. And are those models still valuable?" > > Lin Qiao > > ▶ Watch · 01:01:13

On whether to commit long-term to China for pure data-center speed, given US permitting and legal friction, Qiao granted the construction-velocity gap and complicated it with a labor constraint at home. She has watched a crossover bridge go up in China within a week while a highway near her home remains unfinished after a year, yet even electricians are in severe shortage in the US, which she frames as one more piece of a global supply-chain constraint.

On chips specifically, Qiao explained why she would consider owning data centers before owning silicon. A chip only pays off once the workload is stable, because a taped-out design can't easily change, and AI workloads are still moving too fast.

> "Whenever the usage has passed certain threshold, it makes economic sense to build the underlying supply. But once the hardware is taped out, it's really hard to change it. So once your workload is stabilized, that's the time to consider building a chip. We're still in the early stage of workload maturity to warrant a chip that will be durable." > > Lin Qiao > > ▶ Watch · 01:09:51

Asked for the bottleneck people underrate, she skipped the usual HBM answer and pointed at system design for the largest models. There is still no great system built for very large models, she argued, nothing designed for 10 trillion parameter models today, and closing that gap will take co-design from the model through the serving platform layer all the way to the chip layer, because what needs designing is the system, not any single chip.

Every company should own its intelligence

The concentration worry that runs through the whole conversation lands hardest at the national level. Qiao pointed to a major model being briefly restricted, roughly 19 days and felt most in Europe, as the moment the risk became concrete. Her framing returns to the power line.

> "If we think about the general intelligence model as a power line, every country should own their own power line. It's a very scary moment, my power line is going to be cut off. It's not just whether a country should have their sovereign independence, but every single company should have their independence. You don't want any single person to cut you off." > > Lin Qiao > > ▶ Watch · 01:07:57

The business behind the argument is growing fast enough to make it more than theory. Fireworks is near $800 million of ARR with 200 people, and Qiao expects it to at least double by year end, at a pace Stebbings called unparalleled next to the Slack-era benchmark of reaching $10 million in 18 months.

She closed the leadership section on Jensen Huang with the lesson she took from watching him operate, which is that speed depends on staying close enough to the ground truth to make the call yourself. Leadership, in her telling, is judgment rather than privilege, and in a high-velocity space a leader who does not know what is actually working, yet holds the position to decide, makes the wrong call.

Her single forecast for the next three years ties the strategy and the values into one line, and it's the claim to hold her to.

> "I really see people will own, every single company will own their own intelligence as a must-have. It's not optional. There's an analogy to software. Every single company build their own software stack. Same, I think at a time every single company should own their own intelligence." > > Lin Qiao > > ▶ Watch · 01:27:12

The bottom line

Qiao's argument holds together because every thread runs back to the same question, whether intelligence concentrates into a few general models or distributes into millions of specialized ones. Her token economics, her margin patience, her refusal to move into the application layer, and her warning about one company owning intelligence are the same wager placed at different layers of the stack.

What I find interesting is not the forecast but the stakes she puts under it. She's not arguing about margins or token counts. She's arguing about whether the useful intelligence in the world gets owned by everyone who builds something particular, or by whoever trains the one model the rest of us rent. Both of her futures start from the same place, capable models cheap enough for anyone, and they split only on who does the specializing and who keeps the value. If that specialized layer really is where durable advantage lives, a lot of companies are about to find out whether they hold a belief worth encoding or only a workflow worth renting. She has already chosen the side she's building on. The one she leaves open is whether the rest of the market notices it had a choice before the default quietly becomes the only option.