DeepSeek

In January 2025, a model release from a company most investors had never heard of wiped out roughly $800 billion of global market value in a single day. That was DeepSeek’s R1, and the lesson the market took from it — that frontier AI might not be as scarce, or as expensive, as the incumbents implied — has only sharpened since. It is the most consequential company in this series that almost no Western institution can properly account for, because it refuses to behave like one. No consumer brand to protect, no sales team to feed, no revenue to defend.

DeepSeek isn’t trying to win the AI revenue race — it’s trying to make sure no one can.

The race to zero, made visual

Start with the number that does the damage. DeepSeek prices the model layer at the floor: V4-Flash at $0.14 per million tokens against roughly $5 for a Western flagship like GPT-5.5 or Claude Opus — about 36× cheaper, open-weight, and downloadable to your own hardware. And beneath even that sits the true floor: $0 — run it yourself. After a permanent ~75% price cut, this isn’t a promotion. It’s the strategy.

The economics behind the low price are real engineering, not a stunt. V4-Pro — 1.6 trillion parameters, the largest open-weight model in existence — fires only ~3% of its parameters per token (a “mixture of experts”), runs much of its math in 4-bit precision, and uses a compressed-attention scheme that, at a million tokens of context, needs roughly a quarter of the usual compute. That’s why it can be near-frontier and nearly free.

The number that isn’t there

DeepSeek is private and opaque. It publishes no revenue, its valuation is a press report on a round that hasn’t closed, and even its famous training cost is a single disclosed figure with a large caveat. So before any chart, the grammar — and for DeepSeek, nearly every line lands in the same column.

There is no revenue line to anchor a valuation to — which makes ~$50B both the most speculative mark in AI and the most revealing one. Investors aren’t paying for earnings; they’re paying for a position. Read the rest of this page as the case for why a position with no profit attached can still be worth fifty billion dollars — and why that might even be the correct price.

Commoditize the layer

DeepSeek was founded in July 2023 in Hangzhou as the AI lab of High-Flyer, a quantitative hedge fund run by billionaire Liang Wenfeng, which has financed it entirely — no venture capital until this year, no ads, no enterprise motion. It ships open-weight models under the MIT license (free, downloadable, commercial use permitted) and prices its API at the floor. That posture isn’t idealism; it’s the strategy. A model anyone can run on their own hardware commoditizes the exact thing OpenAI and Anthropic charge a premium for — every download is a small act of erosion against their pricing power, and DeepSeek has no margin of its own to protect while it does the eroding.

In April it sharpened the weapon with V4, its first two-tier lineup — and the tell that this is aimed straight at the incumbents’ wallets is that both V4 models accept OpenAI- and Anthropic-format API calls, positioning DeepSeek as a literal drop-in replacement for either. For the majority of enterprise tasks that never touch the frontier, paying a Western lab ~36× more becomes a line item that’s hard to defend in a budget review.

The fundraise is a control play, not a cash call

For the first time, DeepSeek is taking outside money — and the round says more about the next moves than the balance sheet does. The mark ran from roughly $10B to ~$50B in under six weeks, with China’s state-backed “Big Fund” leading a reported ~$7.35B raise, the largest AI funding round in the country’s history. The purpose isn’t survival — High-Flyer has covered the bills since 2023 — it’s two things the mission now needs: talent equity to compete with US labs poaching staff, and more H20 clusters to scale inference under the export ceiling.

The quiet structural fact underneath: Liang already controls an estimated 84–90% of DeepSeek through High-Flyer, and is reportedly funding roughly 40% of the round himself — so China’s state capital comes in through the front door while his grip barely loosens. A funding round that is really a control consolidation. The one capability still missing is R2, its dedicated reasoning model, stalled for 18 months because the CEO won’t ship below his bar — the main reason DeepSeek is still rated a few months behind the frontier.

The sanction that backfired

The most consequential story about DeepSeek isn’t a model; it’s what US chip controls did to it. Reportedly directed to train V4 on Huawei’s domestic Ascend chips, DeepSeek hit repeated failures — its custom CUDA kernels wouldn’t run on the Huawei architecture — and restarted on NVIDIA’s export-limited H20s, leaving the Huawei silicon for inference only. Read the efficiency tricks again with that in mind and they change meaning: the mixture-of-experts, the 4-bit precision, the compressed attention are, in part, responses to constrained hardware — software wringing frontier results out of second-tier silicon.

Now hold both halves at once. The controls are working in the narrow sense — Brookings judges they’ve delayed China “several years.” But they may be self-defeating in the wider one: by forcing a homegrown stack into existence — Chinese chips already made up ~41% of China’s AI silicon market in 2025 — the same controls are teaching China to build the very capability they were meant to deny. Both true: the sanction bites today, and it compounds an adversary’s independence over a five-year horizon.

V4-Flash, and the week the price floor moved

In late July DeepSeek opened a beta of V4-Flash, an API aimed specifically at agent workloads — the long, repetitive, tool-calling runs where cost per task, not cost per token, is what a buyer feels. It is a beta, not a general release, and the usual caveats about self-reported benchmarks apply. What is not in doubt is the reaction.

Within roughly a week: OpenAI cut API prices on July 30 (its cheap tier by about 80%), Anthropic shipped Opus 5 at about half the prior flagship price, and Google widened its Flash fleet. Four independent companies moved the same lever in the same direction inside one month. That is not coincidence and it is not generosity.

Read it against this piece’s thesis. DeepSeek’s strategic product was never a model — it was a price expectation. The company does not need to win enterprise deals in the West to matter. It only needs to make “good enough, very cheap” a permanent option on the table, so that every rival’s pricing conversation starts from a floor DeepSeek set. On that measure July was its most successful month to date, and it earned no revenue for it.

What argues against it

The valuation rests on undisclosed revenue — the softest mark in this entire series. Open weights are structurally hard to monetize; you can’t meter what people download and run themselves. Tightening H20 controls could block a V5 training run outright, and a sanctioned lab is one policy memo away from a hard ceiling. Compliance is a widening minefield: several governments (Indonesia, Malaysia among them) have already restricted the app on data-sovereignty grounds, and PRC jurisdiction is a non-starter for many Western enterprises before the conversation even begins. DeepSeek may never build a real business. The unsettling part of the bull case is that this doesn’t refute it — it may even confirm it.

My read — the position is the product

I stopped trying to value DeepSeek like a normal company, because it isn’t running the normal race. It’s running a different contest — commoditize the model layer so no US company holds a monopoly on intelligence — and on that mission it has, by any honest scoring, already won. V4-Flash prices frontier-adjacent capability at near-zero, tens of millions of downloads seed a global open-weight ecosystem under Chinese stewardship, and the fundraise just bought it the compute and the talent to keep pressing.

The number I’m watching is V5 under H20 constraints.

If a sanctioned lab on second-tier chips reaches genuine parity with GPT-5.5 and Claude Opus 4.8, the closed-model pricing power of the entire Western industry comes into question in a single release — and the bifurcation Brookings describes, a US stack and a China stack running in parallel, becomes the defining fact of the decade rather than a forecast.

If I’m building: for the large share of work that doesn’t need the absolute frontier — routine extraction, classification, drafting, especially where data must stay on-premises — an open-weight V4-Flash deployment is the cost-control option I’d at least price out before signing a Western API contract. The catch is governance and jurisdiction, and for regulated data that catch is often disqualifying. But ignoring the floor it sets would be a mistake.

Sources: valuation and the ~$7.35B raise (China’s Big Fund leading; Tencent, CATL in talks), Reuters and WSJ; the raise target and Liang’s control consolidation, The Information; V4 specs, two-tier lineup, and OpenAI/Anthropic-format compatibility, DeepSeek docs and CNBC; the permanent ~75% price cut, Reuters; the Huawei/H20 hardware story, OSINT report; export-control and bifurcation analysis, Brookings; R2 delay, Reuters. The honest frame: this is a funding/valuation story, not a revenue one — read every number as reported, not confirmed, and the blank revenue line as the truest figure on the page.

This deep dive is part of “Deep Research on Frontier Companies.” Crafted with intent by Ravi Teja Palanki · June 2026.