By Matt Busel, who built the apps, crates and game he writes about · every number links to its source and every dataset is downloadable (methods) · corrections · featured in Kirino, gamebiz, This Week in Rust, New Mobile Life, By-Press
DeepSeek's $15B round: 5% of OpenAI's price, 22% of its API bill
On this page · 8 sections
Short answer: DeepSeek closed about 100 billion yuan, roughly $15 billion, from China's industrial giants led by CATL and Tencent, two days after the US ordered every civilian federal agency to rip DeepSeek out. It asked for a valuation of about $75 billion, around 5% of the $1.4 trillion OpenAI is pitching, while its flagship API lists at 22% of OpenAI's at peak hours and 11% off-peak. Washington banned the cheap option; Chinese industry bought it at a discount.
TL;DR: The round grew from a 50 billion yuan target to 80 billion and closed near 100 billion yuan ($14.96 billion at the Fed's rate), per Bloomberg on Oct 22, two days after a DHS order published in the Federal Register on Oct 20. One million tokens in and one million out costs $5.28 on DeepSeek V4-Pro at peak against $24 on GPT-5.6 Sol or Claude Opus 5.5 and $60 on Claude Fable 5.1. DeepSeek's open models pulled 24.4 million Hugging Face downloads in the last 30 days, more than Meta's Llama org at 16.0 million, and at least 4,857 models are built directly on just 12 of its releases, all now barred from civilian federal IT.
Key numbers
| What | Number | Source, checked Oct 25 2026 |
|---|---|---|
| DeepSeek round size | ~100bn yuan = $14.96B | Bloomberg via TNW, FRED DEXCHUS 6.6861 |
| Valuation sought | ~500bn yuan = $74.8B | CNBC Oct 6, via CellCog |
| OpenAI: last round / talks | $852B post (Mar) / ~$1.4T (Sep 30) | Bloomberg via ProPakistani |
| Anthropic Series H valuation | $965B (May 2026) | Trending Topics / FT |
| DeepSeek V4-Pro list price, peak | $1.32 in / $3.96 out per 1M | DeepSeek pricing page |
| GPT-5.6 Sol / Claude Opus 5.5 | $4 in / $20 out per 1M | OpenRouter models API |
| DeepSeek HF downloads, 30 days | 24.4M (106 models) | Hugging Face API |
| Direct derivatives of 12 DeepSeek models | 4,857 | Hugging Face API, base_model filter, our pull |
| DHS exclusion and removal order | Oct 20 2026, 91 FR 66346 | Federal Register |
The ban came Tuesday. The check came Thursday
On Tuesday Oct 20 the Department of Homeland Security published an order under the Federal Acquisition Supply Chain Security Act. It excludes DeepSeek products from civilian federal procurement and orders agencies to remove them, including "third-party products and services developed using DeepSeek implementations."
On Thursday Oct 22, Bloomberg reported the round had closed near 100 billion yuan. DeepSeek first aimed for 50 billion, raised the target to 80 billion when demand ran hot, and ended at about 100 billion, per The Tech Portal's summary of the Bloomberg report. DeepSeek, CATL and Tencent have not confirmed it, and Reuters says the details could still change.
That sequence is the story. The ban did not scare the money. It doubled the order book.
Priced like a mid-cap, built like a frontier lab
DeepSeek sought about 500 billion yuan, or $74.8 billion at the Fed's Oct 16 rate. OpenAI raised $122 billion in March at an $852 billion post-money valuation and is now pitching about $1.4 trillion, per Bloomberg. Anthropic's May Series H set $965 billion.
So DeepSeek costs 5.3% of OpenAI's ask, 7.8% of Anthropic's last mark and 8.8% of OpenAI's March price. Bloomberg's own framing is that OpenAI and Anthropic have each raised more than $100 billion, so DeepSeek's $15 billion is still the smaller war chest. Our read: a smaller war chest is the point. DeepSeek has spent two years proving it does not need one.
The climb inside 2026 is steep but not absurd: about $50 billion at the June round (more than $7 billion raised), a reported $71 billion ask in July, about $75 billion now. That is 50% in four months. OpenAI's ask went from $852 billion to $1.4 trillion in six months, a 64% jump on a base 17 times larger.
The API bill: 22% at peak, 11% off-peak
Valuation is a story. Price per token is a receipt. We pulled DeepSeek's official pricing page and OpenRouter's public model list on Oct 25 and priced one million input tokens plus one million output tokens.
- DeepSeek V4-Pro, peak: $1.32 in, $3.96 out, $5.28 total.
- DeepSeek V4-Pro, off-peak (all weekend, plus most weekday hours): $2.64 total.
- DeepSeek Flash, peak: $1.50 total. Cached input on Flash costs $0.006 per million.
- OpenAI GPT-5.6 Sol and Claude Opus 5.5: $24. GPT-5.5: $35. Claude Fable 5.1: $60.
DeepSeek's flagship runs at 22% of GPT-5.6 Sol at peak and 11% off-peak. Against Fable 5.1 it is 8.8% and 4.4%. Off-peak means every hour outside 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, so a US team running overnight batch jobs pays the half rate almost by default.
One change to watch: DeepSeek's page says its Search and Fetch tools, free today inside agent harnesses like Claude Code and Codex, start billing at $0.75 per 1,000 requests on Nov 1. Cheap labs still find the meter. They just set it lower.
Follow the money: who wins, who loses
CATL and Tencent win an option. A battery maker and a messaging giant get a frontier lab at 5% of the American price, with a stock market listing reportedly in preparation. Bloomberg also reports DeepSeek's cloud already runs on Huawei chips, so the capex stays inside China's own supply chain.
Their shareholders have not cheered yet. From Sep 24 to Oct 22, CATL fell 24.9% (293.50 to 220.41 yuan) and Tencent fell 12.8% (HK$438.40 to HK$382.40), then both bounced on Oct 23. Nvidia fell 6.1% over the same stretch. We are not claiming the round caused it; a broad Asia tech selloff ran the same weeks. We are saying the people writing the check bought in a drawdown.
US labs keep their federal moat. The order hands OpenAI, Anthropic and Google the civilian federal market without having to win it on price. Taxpayers fund the gap: every agency workload that could have run at $5.28 per million-pair now runs at $24 or more on an approved vendor. Our bet is that the "national security" framing does most of its work on the procurement side, not the security side, because open weights do not phone home when you run them on your own hardware.
Open-source builders are the collateral. DeepSeek's 106 Hugging Face models pulled 24.4 million downloads in 30 days and 409.9 million all-time. Meta's Llama org pulled 16.0 million in the same 30 days, Mistral 9.0 million, Moonshot's Kimi 2.4 million. Only Alibaba's Qwen (245.7 million) and the big multi-purpose orgs out-pull it.
The order's "third-party products developed using DeepSeek implementations" line reaches into that ecosystem. DeepSeek-R1 alone has 524 direct derivatives; the tiny R1-Distill-Qwen-1.5B has 1,446. Across just 12 DeepSeek releases we count 4,857 models that name one as their base. Any contractor shipping one of those into a civilian agency now has a compliance problem.
What it means for the AI price war
Two pricing systems now run side by side. In one, a lab raises at $1.4 trillion and charges $20 per million output tokens. In the other, a lab raises at $75 billion and charges $3.96, or $1.98 on a weekend. Both claim frontier quality, and our Qwen3.8 vs Fable 5 repricing showed the benchmark gap is narrower than the price gap.
The US answer so far is procurement bans, not lower prices. That protects incumbents for exactly one customer, the federal government. Everyone else, from a two-person startup to a Fortune 500 data team, can still route to whichever endpoint is cheapest. Capital just told you which side it thinks wins that fight: $15 billion in two weeks, double the original target.
For more on what these models cost to run, see our LLM price hub, how DeepSeek gave away its inference engine, and the paper trail on Anthropic's pre-IPO SPVs. Every deep tech and money story lives on Tensorust News and the deep tech and money hub.
What this means for you
If you sell to the US federal civilian government, audit your stack this week for any DeepSeek model or fine-tune, including distills buried inside a vendor product; the order names third-party products too. If you do not, rerun your monthly AI bill at DeepSeek's off-peak rate and at your current vendor's, and decide with the real gap in front of you, not the headline.
Our $7 LLM API Cost Calculator does that math per request, per month and per year across 26 models. If it does not give you a clear monthly cost, we keep fixing it at no extra cost until it does what it promises.
Not financial advice. Our numbers and our read, so you can check them. Download the data below.
Sources
- Federal Register, DHS Notice of Order on DeepSeek, 91 FR 66346, docket DHS-2026-0628, Oct 20 2026
- The Next Web, DeepSeek reportedly raises $15bn from Tencent and CATL ahead of IPO (citing Bloomberg), Oct 23 2026
- The Tech Portal, DeepSeek raises $15 billion, doubling its own target, Oct 22 2026
- CellCog, DeepSeek Raises $15 Billion: What Is Reported, What Isn't (CNBC Oct 6 valuation), Oct 22 2026
- ProPakistani, OpenAI reportedly seeks $30 billion at $1.4 trillion valuation (citing Bloomberg), Sep 30 2026
- Trending Topics, Anthropic investors expect IPO valuation of $2 trillion (citing the FT), Aug 13 2026
- DeepSeek API, Models and Pricing, checked Oct 25 2026
- OpenRouter, models API (list prices), our pull Oct 25 2026
- Hugging Face, model API (downloads, base_model derivatives), our pull Oct 25 2026
- FRED, DEXCHUS China / US exchange rate, Oct 16 2026 value
- Yahoo Finance, daily closes for 0700.HK, 300750.SZ, NVDA, our pull Oct 25 2026
First published here on tensorust.com on Oct 25, 2026. Related: Tensorust News, Deep tech and money, LLM price hub, DeepSeek gave away its engine, Qwen3.8-27B vs Fable 5 price, Anthropic IPO: 79 SPVs.
Up next
More on Deep tech and moneySpaceX's $40B Nvidia chip debt costs 8.3%. Apollo is the lenderSpaceX paid 5.35% in June. Its $40B Nvidia chip debt reportedly prices at 7.75-8.3%, above the junk index, costing $3.3B a year vs $1.7B of AI profit.Read it next →Also worth your time
Cite this
Free to cite and reuse under CC BY 4.0 with a link back. How it was collected: methods. Spotted an error? Corrections.
Did this answer your question?
Get the next one by email.
What my AI agents tried this week, with the real numbers: what worked, what flopped, what made money. One email when there is something worth reading.
No spam, never sold. Every email has a one-click unsubscribe.
What are people saying about this? Read and post on The Floor →
More: the author, related reading, share this, a free game
Try it yourself · Want the numbers to feel real?
See any amount as a pile of cash
Type an amount, see it as $100 bills next to a person: weight, volume and height.
In The Manual: on this topic
Explainers answer one question, short answer first. Guides walk you through it step by step.
- Explainer · Money and the economyHow Kalshi weather markets work
- Explainer · Money and the economyCan the Fed buy Treasuries directly from the government?
- Explainer · Money and the economyThe Fed's deferred asset, explained
- Explainer · Money and the economyWhat is SOFR?
Read next
- Economics · Oct 25, 2026 · 9 min readSpaceX's $40B Nvidia chip debt costs 8.3%. Apollo is the lender
- Economics · Oct 25, 2026 · 9 min readGPU depreciation: Big Tech says 6 years, a 2020 A100 rents for $0.74
- Economics · Oct 25, 2026 · 8 min readNvidia's $2,000-per-GPU software fee: rent on a stack vLLM gives away