Fireworks AI's $40 Billion Ask Values the Inference Software Layer at 40x Revenue — and NVIDIA, Its Supplier and Shareholder, at 17x

Fireworks has marked itself up three times in twelve months, and the third step is still only a conversation. Each step up was funded by a fresh primary round, with no profitability disclosed, which is why the October ask reads as a financing signal rather than an earnings signal.

The load-bearing uncertainty is the denominator. Fireworks has not published a revenue figure since July 15, so the 30–40x multiple is computed on a run rate that is at least three months stale. If the run rate has compounded since July at anything like the pace it did through the first half of 2026, the effective multiple is meaningfully lower than the headline.

A 30–40x revenue mark only makes sense if the business underneath behaves like software. Fireworks' does not. The estimated 50% gross margin at Fireworks sits below the 65.9% CoreWeave earned on rented GPUs in its June quarter, which means the layer collecting the software multiple buys its compute on worse terms than a bare-metal neocloud does.

The multiple table hides the part that matters. The inference software layer does not make chips; it buys them, and it buys them with the money it raises. Fireworks' $1.505 billion Series D was explicitly earmarked for engineering and global compute capacity, so every dollar of private equity that lands in the software layer reappears as chip orders one step down.

So is NVIDIA's inference pricing power being repriced away? Not yet, and its own margin line says so. Blended gross margin expanded to 75.0% in the June quarter, up 2.6 points year over year, even as the software layer's marks doubled. What is being repriced is the financing of chip demand, not the price of chips: the marginal buyer of inference GPUs is now a software company whose equity value rests on a multiple with no cash-flow floor beneath it.

The layer being marked highest is the layer that owns the least risk. The layer actually buying the silicon carries it on debt, and the software layer is increasingly the buyer's first call. CoreWeave's $104.2 billion backlog and $35–39 billion of 2026 capex are financed at 11.5 times net debt to EBITDA, while its enterprise value of $79.5 billion is only 10.5 times trailing revenue.

The other end of the chain is where the value lands. Inference pricing is a cost line for everyone building on top of it, and the October repricing only matters if it lowers that cost. A layer marked at 40x revenue is pricing on the assumption that a token will get materially cheaper — that is what these marks actually underwrite.

In the next two quarters, three things resolve this. NVIDIA reports third-quarter fiscal 2027 results on Nov 17, 2026 against a $108 billion revenue outlook, CoreWeave prints a third quarter with $3.4–$3.6 billion of guided revenue and $35–39 billion of capex, and Fireworks either lands a lead investor or does not. Only the third of those tests the thesis; the other two test the compute layer underneath it.

Over one to three years, the interesting question is where inference margin settles. If Fireworks' 60% gross-margin target is real, the software layer takes margin from the neoclouds and the chip makers keep their pricing power. If it is not, Fireworks stays a GPU reseller marked like a software company, and the gap closes the other way — token prices fall until hardware margins compress. The second outcome is the one Etched's $21 billion rack-scale bet and Tenstorrent's $6.2 billion round are financing.