The one-screen version: in the week of October 20–23, three of the largest capital events in the history of this industry stacked almost on top of each other. DeepSeek reportedly raised ~$15B (≈100 billion yuan, led by CATL and Tencent) at a reported ~$74B valuation. OpenAI is racing to close up to ~$30B at a $1.25 trillion pre-money valuation. And Anthropic is lining up a mid-November IPO reportedly seeking $1.5–2 trillion. For a founder building on these companies, the headline isn't the scoreboard. It's that none of this money is raised against profit — it's raised to pre-buy compute — which makes the price you pay for inference an investor subsidy, not a market rate.

The three events, briefly#

DeepSeek, ~$15B, and a wall. Per TechCrunch and Bloomberg, DeepSeek pulled in roughly 100 billion yuan from more than 15 investors, with battery giant CATL and Tencent among the largest, at a reported valuation near $74B. Coverage disagrees on whether the round has fully closed, so treat the number as reported, not settled. The timing is the part to notice: it landed days after the US ordered federal civilian agencies to strip DeepSeek software from their systems. A frontier open-weight lab just got funded for years and fenced off from a major market, in the same week.

OpenAI, ~$30B, against the clock. CNBC reported OpenAI nearing a round of up to $30B at a $1.25 trillion pre-money valuation, with SoftBank and Abu Dhabi's MGX in talks and a stated goal of collecting signatures within about a week. Earlier Reuters reporting put the target higher, above $1.4T — the moving figure is itself a tell that terms are unsettled. The framing in every outlet is the same: OpenAI wants this done before Anthropic's IPO resets what "a frontier lab" is worth.

Anthropic, mid-November, and a $600B tab. Anthropic is reportedly targeting a $1.5–2T listing. The detail that matters for anyone building on it is in the draft prospectus the WSJ reviewed: an $80.7B net loss paired with more than $600B in future compute commitments, with a second ~$45B Nscale compute deal reportedly in preparation. That is not a company with a pricing problem. It's a company making an enormous bet on demand that hasn't fully arrived.

Read the three together and the shape is obvious: this is not a profit race, it's a compute-procurement race. The capital is being raised to buy GPUs and power, and the per-token price is the loss leader that fills them.

The one real idea: your API price is a promotional rate#

Here is the part that transfers straight into your architecture. When a provider raises tens of billions against future revenue and commits hundreds of billions to compute it hasn't filled yet, it has every incentive to price inference below its fully-loaded cost to win developers now. That's not cynicism — it's the standard logic of a land-grab, and the prospectus numbers make it legible. The cheap frontier tier you're building margins on is, in part, underwritten by investors betting on a 2028 that may or may not show up.

For a solo founder or a small team, that has one concrete consequence: the model endpoint is the least stable input in your whole cost structure, and it's the one you're most tempted to hard-code. When the funding race normalizes — when a round doesn't close, or an IPO reprices the sector, or committed compute has to start paying for itself — the first thing that moves is the price of the cheapest tokens. You don't want to discover your margins depend on a promotional rate the same week it ends.

What to do before the next megaround#

The move isn't to predict the winner. It's to stop needing to.

The week's three megarounds will be remembered as a market peak or a market top; nobody knows which yet. Either way, the lesson for someone building is the same one it always is when other people are betting trillions on your inputs: own the seam, keep an exit, and don't mistake a subsidy for a price.