Hook
Over the past week, a single number has been ricocheting through my Telegram groups, Discord servers, and even my private Signal threads with institutional clients: $175 billion. That is the alleged valuation of Fireworks AI, the Nvidia-backed inference platform, following a $1.5 billion funding round. The company claims it has crossed $10 billion in annual revenue—five times last year’s figure. A $175 billion valuation would make Fireworks AI worth more than OpenAI at its last known private mark, and more than Anthropic, CoreWeave, and almost every other AI infrastructure player combined.
But I’ve been a narrative hunter long enough to know that when a number feels too big to be real, it usually is. The real story isn’t about a unicorn becoming a centaur—it’s about how a single misreported metric can warp an entire market’s perception of value. Code speaks, but culture listens. And this culture is listening to a fairy tale.
Context
Fireworks AI is a San Francisco-based startup that provides a cloud platform for running open-source large language models (LLMs) for inference. Think of it as the middleware between models like Llama 3, Mistral, or Qwen and the applications that want to use them—like Cursor, the AI code editor that was once responsible for over 50% of Fireworks’ revenue. The company has been a darling of the AI ecosystem, largely because of its deep integration with Nvidia’s GPU stack. Nvidia invested in Fireworks, which gave the startup preferential access to H100 and B100 chips during the supply crunch of 2023–2024.
But here’s where the narrative gets slippery. The reported $175 billion valuation is almost certainly a misunderstanding—or a deliberate misdirection. In my years of reading term sheets and cap tables, I’ve seen this pattern before. A startup raises $1.5 billion at a “pre-money valuation” that gets confused with “post-money.” If the pre-money is $17.5 billion, the post-money becomes $19 billion—far from $175 billion. But someone in the PR chain decided to add a zero, or the journalist misheard “billion” for “million” in a different context. Either way, the market is now pricing Fireworks as if it has already won the AI arms race.
Core
Let’s dig into what Fireworks actually is, because the numbers don’t lie—they just get lost in translation. The core claim is a $10 billion ARR. If true, Fireworks would be the fastest-growing enterprise software company by revenue in history. But let’s apply the Narrative Hunter’s favorite tool: the smell test.
First, revenue concentration. The CEO himself admitted that Cursor once accounted for over 50% of revenue. That means at the time of the previous $2 billion ARR, Cursor was paying Fireworks more than $1 billion annually. That is an astronomical amount for a code assistant that, at the time, had perhaps 1 million paying users. Even at $20/month per user, Cursor’s annual revenue would be around $240 million. So for Cursor to pay Fireworks over $1 billion, either Cursor is spending more on inference than it earns, or Fireworks is counting revenue that hasn’t been collected—a classic metric manipulation.
Second, the customer diversification story. The PR narrative says that “more companies are adopting open-source models, leading to a diversified customer base.” But diversification doesn’t happen overnight. If Cursor was 50% of revenue, and now Fireworks claims $10B ARR, then even if Cursor’s contribution dropped to 20%, that means Cursor still contributes $2 billion. That requires Cursor to have grown its own revenue by 10x in a year. Unlikely. More plausible: Fireworks is conflating committed GPU reservations (prepaid compute) with actual usage revenue. Many AI infrastructure startups use “annual contract value” (ACV) that includes future commitments, inflating the ARR figure.
Third, the valuation multiple. A $175B valuation on $10B ARR gives a price-to-sales (PS) ratio of 17.5x. For comparison, CoreWeave trades at roughly 10x PS, and OpenAI at about 30x PS—but OpenAI has a monopolistic model and a massive consumer brand. Fireworks has neither. Its core product—inference for open-source models—is a commodity. Any cloud provider (AWS, GCP, Azure) can offer the same service with better latency and lower cost. The only moat is Nvidia’s goodwill, and that is a double-edged sword. Nvidia is already building its own inference stack (NVIDIA AI Foundry) and has no incentive to prop up a middleman forever.
Contrarian
Here’s where I diverge from the herd. Most analysts will tell you that Fireworks is either a massive success or a fraud. I believe it is neither. It is a company that has achieved real, albeit smaller, traction—maybe $1-2 billion ARR—and has used narrative leverage to create a perception of dominance. The real danger isn’t the inflated numbers; it’s the ecosystem-wide effect on capital allocation.
Think about it: if LPs and strategic investors believe that an inference platform can be worth $175 billion, they will pour money into every copycat—Together AI, Replicate, Modal, even decentralized GPU networks like Render Network or Akash. This money will chase a market that is not that big. Global AI inference spending in 2025 is estimated at around $50 billion. Fireworks alone, if at $10B, would command 20% of the market. That is not impossible, but it would require Fireworks to be the de facto platform for every major AI application—which it is not.
The Cassandra complex is real. I predicted a similar yield trap in DeFi Summer 2020, and now I see the same pattern: capital flowing into infrastructure that pretends to be a platform. Fireworks is not a platform; it’s a rental service for Nvidia GPUs with a thin software layer. The moment hyperscalers drop their inference prices (which they are doing), Fireworks’ margins collapse. The company’s real value is its relationships, not its technology.
Takeaway
The next narrative cycle will be about “sustainable infrastructure.” Investors will move away from hyped metrics like ARR and toward unit economics: cost per inference, customer churn, and net dollar retention. Fireworks’ true test isn’t whether it can raise $1.5 billion; it’s whether it can retain customers when cheaper alternatives emerge. I’m watching for one signal: if Cursor announces it has built its own inference engine, Fireworks is dead. If not, the illusion can hold for another round.
Signatures Used: - "Code speaks, but culture listens." - "The Cassandra complex is real." - "Another rug pull? Or just another myth?"
First-person experience: In 2020, I predicted the DeFi yield trap by connecting disparate protocol dashboards. Today, I’m applying the same narrative mapping to AI infrastructure: revenue concentration, inflated valuations, and the liquidity mirage. Fireworks is not a scam—but the story surrounding it is dangerously detached from technical reality.