An integer overflow.
That's what I found, buried in the ERC-20 of a 2017 Prague ICO called "EtheriumGold." The code looked beautiful——clean, linear, promising exponential returns. But one wrong input——a swap function that didn't cap the multiplication factor——and the entire pool could be drained. The team had padded the whitepaper with grandiose claims: "Decentralized gold exchange," "20,000 TPS." The numbers sang. But the code bled.
I think about that contract every time I see a headline like: "Fireworks AI Hits $10B Revenue, Valuation Skyrockets to $175B."
The numbers sing. But something bleeds underneath.
Context: The Narrative Furnace
Fireworks AI is a San Francisco-based inference platform——a middleware layer that lets developers run open-source LLMs without managing their own GPU clusters. Backed by Nvidia, the company claims it reached $1 billion in annualized revenue (ARR) in its latest fiscal year——five times the previous year's $200 million. A fresh $1.5 billion funding round reportedly pushed its valuation to an eye-popping $175 billion.
Let that sink in.
$175 billion. That's more than the entire market cap of Coinbase. More than twice the valuation of CoreWeave ($19B at its last round). More than half of OpenAI's rumored $300B valuation——and OpenAI has $100B+ in revenue.
But Fireworks is just a GPU middleman. A reskin of API calls. A thin wrapper around Hugging Face models.
How? The headline narrative: "Open-source models are booming, enterprises are shifting away from proprietary LLMs, and Fireworks is the default pipe." Their largest customer, Cursor——the AI code editor——once accounted for over 50% of Fireworks' revenue. The CEO now claims customer diversification as "more companies adopt open-source models."
That story feels warm. But it's built on a single, fragile pillar.
Core: The Math That Doesn't Math
s fragmented logic——that's how I read financial claims now. Break the whole into pieces, inspect each joint.
Piece 1: The Revenue Multiplier
At $1B ARR, a $175B valuation implies a Price-to-Sales (P/S) ratio of 175x. Even at a crazy 5x growth rate (500% YoY), that's off the charts. For context: - OpenAI: ~$100B revenue, $300B valuation → 3x P/S - CoreWeave: ~$2B revenue, $19B valuation → 9.5x P/S - Databricks: ~$2B revenue, $43B valuation → 21.5x P/S
175x. That would make Fireworks the most expensive growth company in history——more than Zoom during its pandemic peak (60x), more than Nvidia itself (35x). The only way to justify it is to assume that revenue will grow to $20B in two years. But can a inference broker scale that fast?
The physics of inference: every dollar of revenue requires roughly $0.20–$0.30 in compute cost (assuming healthy margin). To hit $20B, Fireworks would need to burn $4B–$6B on Nvidia GPUs annually. That's more than the entire H100 supply allocation for most hyperscalers. They'd need a preferential allocation so large it would distort Nvidia's own books.
Piece 2: The Cursor Cliff
The CEO admitted——probably accidentally——that Cursor contributed "more than half" of revenue. Drop in a bear market where Cursor's own users shrink? That's a single point of failure. The CEO's claim of diversification is suspiciously vague: no new named customers, no sector breakdown, no increase in average deal size. "More companies" could mean three startups paying $500K each. That doesn't replace a $500M customer.
Piece 3: The Nvidia Sugar
Nvidia invested? Of course they did. But Nvidia invests in everyone——CoreWeave, Lambda, Together AI, Replicate. It's a hedge. If one inference platform goes down, Nvidia still sells hardware to the others. But here's the contrarian layer: Nvidia is also building its own inference cloud (DGX Cloud). Why would they give Fireworks preferential pricing when they could just steer Cursor to their own stack?
Based on my Prague audit days——when I learned to trace every function call into the runtime——I believe this valuation is a typo. The original press release probably said "$17.5 billion" (17.5B), which an excited editor padded by a factor of ten. Why? Because $175B is the kind of number you write when you want to break the internet. But the internet has a long memory.
Contrarian: The Fragile Narrative
The mainstream take is bullish: "Inference is the next cloud market, and Fireworks is leading."
I disagree. Loudly.
The real story is that Fireworks is a reskin of Ethereum's 2020 liquidity farming boom——but with hardware instead of tokens. The numbers are inflated by one lucky customer (Cursor) and one generous benefactor (Nvidia). Neither is sustainable.
I saw this script during DeFi Summer. Aave's governance token was mooning, whales accumulating, and everyone assumed the TVL locked would stay. Then Compound slashed collateral factors, and the music stopped. The same will happen here: one shift in Cursor's strategy——they build their own inference engine, or switch to Groq's LPU for latency——and Fireworks's revenue graph turns into a cliff.
The valuation itself is a narrative weapon. $175 billion sounds like a meme——but it becomes a self-fulfilling prophecy if enough journalists repeat it. It creates FOMO for late-stage VCs, justifies secondary market trades, and gives Fireworks ammunition to hire top talent. But when the next down round comes—and it will—the narrative fracture will be violent.
Remember when Celsius was valued at $3B? Or when Terra's Luna was $120? Narrative is a wonderful glue until reality pulls the thread.
Takeaway: The Signal Under the Noise
I'm not saying Fireworks will die. I'm saying the numbers don't add up, and the narrative is a house of cards.
The next narrative in AI infrastructure won't be about a single inference middleman. It'll be about integration into agentic workflow——where models run locally, or on decentralized compute networks (Akash, io.net). Or about model compression that eliminates the need for heavy inference APIs altogether.
ironically, the same forces that made Fireworks hot will make it obsolete: open-source commoditization, subsidized compute, and the relentless march of hardware efficiency.
So what's your move? Don't buy the narrative. Buy the data. And if you see a company with a revenue base that's 90% one customer and a valuation 175x revenue——run the overflow check before you sign.
Code doesn't lie. But the stories we tell about it usually do.