DeepSeek's Price Hike Resets the AI Infrastructure Investment Thesis

DeepSeek announced a price hike today. The market chatter is predictable: the race-to-the-bottom is over, pricing power is returning, and the AI infrastructure investment thesis just got safer. That interpretation is wrong — and it matters for anyone holding Nvidia, the hyperscalers, or data center REITs.

Let me be precise about what DeepSeek actually did. DeepSeek announced a 2x price increase for its API services. Off-peak rates remain unchanged.

The current base rates are $0.14 per million input tokens for V4 Flash and $0.435 for V4 Pro. Even at the projected 2x rate — $0.28 for Flash input, $0.56 for Flash output — DeepSeek remains the price floor of the frontier-adjacent API market by a wide margin. The 100x gap the market fixated on last year was always a comparison between DeepSeek's cheapest and the most expensive Western flagship. The typical comparison is about 27x. After this hike, it is still about 27x.

The 100x gap is not closing. It is barely shrinking.

Now track what the Western labs are doing — because that is where the real pricing signal lives. OpenAI is reportedly considering "drastic" further token-price cuts to compete with Anthropic for enterprise clients.

Why raise prices at all, then? Two reasons, neither of which resets the infrastructure thesis. First, capacity management. The peak windows are when demand concentrates — the surcharge is a throttle, not a strategic repricing. Second, IPO preparation. DeepSeek is reportedly planning to go public, and improving unit economics ahead of a listing is standard behavior. Every lab spending billions on compute wants to show a path to profitability. That does not mean the competitive dynamic has flipped.

Now the question that matters for capital allocation: does this change the AI infrastructure demand thesis?

The thesis for Nvidia, the hyperscalers, and the data center REITs has never rested on whether DeepSeek charges $0.14 or $0.28 per million tokens. It rests on total compute demand — the volume of tokens being generated, the number of applications being deployed, the pace of enterprise adoption, and the structural buildout of capacity to serve it. Lower per-token pricing is a feature, not a bug, for the infrastructure thesis. It drives more users, more applications, more tokens, and more demand for compute.

Nvidia just reported 70.68% revenue growth, a 64% operating margin, and an 89% return on invested capital. Its inventory sits at $21.4 billion — not a sign of excess, but a signal of supply being built to meet demand that is still accelerating. The hyperscalers are locked in a capacity arms race: Microsoft, Amazon, and Google are competing for AI workloads, and every one of them has increased capex guidance this year. Data center REITs like Digital Realty are up 24.7% year-to-date, reflecting the demand for physical infrastructure.

The risk that would break the thesis is not that DeepSeek doubles its price. It is that compute demand decelerates. That is not happening.

The debate is not whether DeepSeek's hike signals a return to pricing power across the AI stack. It is whether the opportunity cost of holding infrastructure names has changed. I do not believe it has. Nvidia at $5.28 trillion market cap, with a PEG ratio of 0.30x and a trailing P/S of 20.8x, is still priced for the growth it is delivering. The ~70% revenue growth rate exceeds the multiple. The hyperscalers are trading at levels that reflect their cloud-and-AI duopoly positions. The data center REITs are pricing in the capacity buildout — and they should be.

If I were looking for a reason to reduce my allocation to AI infrastructure, a 2x price hike from the market's cheapest provider — who remains the market's cheapest provider after the hike — would not be it. The 100x gap narrative was always more headline than reality. The roughly 27x gap that actually exists is intact. And the structural demand that makes the infrastructure thesis work is still accelerating.

The thesis does not change. The allocation stays.