Washington bans DeepSeek, DeepSeek gives away its engine, and Amazon's 'dark GPUs'

The US Department of Homeland Security (DHS) has ordered civilian federal agencies to stop buying products and services from Hangzhou DeepSeek Artificial Intelligence and to remove DeepSeek software already running on their systems. CyberScoop and Engadget reported the order on October 19, with the notice scheduled to appear in the Federal Register on Tuesday, October 20. According to CyberScoop, the letter was signed by Homeland Security Secretary Markwayne Mullin months earlier, meaning the ban has quietly been in force for some time and is only now public. Engadget reports that DeepSeek is also barred from federal contracts.

CyberScoop says the order relies on the federal supply-chain law that lets the government exclude risky technology, following a review by the Federal Acquisition Security Council. ITPro reports that the scope goes beyond DeepSeek's own apps: it covers "third-party products and services developed using its implementations" and applies to prime contractors and subcontractors serving agencies such as Treasury, State and Justice. Agencies that used DeepSeek are told to check whether they were compromised, monitor traffic and add logging. Agencies can ask for exceptions with a "compelling justification", but CyberScoop says requests must be filed 60 days in advance.

The stated reasons are data and malware risk. In the notice, DHS says DeepSeek stores data on servers in China, that its hosted service collects personal information when users create accounts, and that prompts and responses "may result in the transfer of sensitive data" to the Chinese government. Engadget quotes the notice as saying the models could "contain malware or other security vulnerabilities." ITPro notes DeepSeek is already restricted in Texas, New York, Virginia, Australia, South Korea and Taiwan, and by the US Commerce Department and Navy. The order landed days after DeepSeek released its V4.1-Flash-1017 model.

Several important questions are unanswered. Reports do not agree on which products count as "developed using" DeepSeek, and none of the coverage explains how agencies should treat DeepSeek's open-weight models when they run on a US company's servers with no data going to China. No removal deadline has been reported. DeepSeek had not publicly responded in the coverage we reviewed, and CyberScoop said DHS and CISA had not answered its questions. Readers should treat the contractor and third-party details as reported by ITPro until the full notice text is widely analysed.

The practical effect is that model provenance is becoming a compliance question. If you sell software, consulting or AI features to US federal agencies, or to companies that do, buyers will start asking which models are inside your product, where they run and who can see the data. That creates work for anyone who can inventory AI use, swap models without breaking products and document the result. It also sends a signal to other governments and regulated industries that already watch US procurement rules, including many in Africa that buy from US-based vendors.

Why it mattersThis moves AI model choice from an engineering preference to a procurement and compliance issue, and the reported reach into contractors and third-party tools means the effect spreads well beyond government offices.

On Tuesday, October 20, DeepSeek published DeepInfer on GitHub and announced it on X. The README describes it as "the high-performance inference framework that powers DeepSeek's large-scale production services." Inference is the work of actually running a trained model to answer requests, and it is where most of the ongoing cost of an AI service sits. DeepInfer is released under Apache 2.0, which allows commercial use and modification. The Next Web reported more than 8,000 GitHub stars by Tuesday afternoon; when we checked, the repository showed about 12,000 stars and 1,100 forks.

For now DeepInfer supports a single model, DeepSeek-V4.1-Flash, and two kinds of hardware: Nvidia SM100-class data-centre GPUs and Huawei's Ascend 950 chips. It needs Linux, Python 3.12, and CUDA 13 on Nvidia or CANN 9.2 on Ascend. Features include custom kernels, ways to split work across many chips, speculative decoding (drafting up to five tokens ahead by default and checking them in one pass), asynchronous scheduling, and structured output that forces answers into a regex, JSON schema or grammar. Deterministic, repeatable outputs are available on Nvidia only.

The README says DeepInfer was built mainly on vLLM, the popular open-source serving engine, with inspiration from SGLang and Nvidia's TensorRT-LLM. The Next Web reports that vLLM thanked DeepSeek and said it would improve its own V4.1-Flash support using ideas from DeepInfer. DeepSeek said it is sharing its production stack "to help the community lower compute costs, reduce energy use, and make AI more affordable." It also released deepseek-recipe, a tool that matches how its official API turns text into tokens. The release came a day after the US order banning DeepSeek from civilian federal agencies.

There are clear limits. This release covers offline inference and leaves out three things a production service usually needs: a KV cache system, separation of prompt processing from answer generation, and load balancing across replicas. The README says most of the engine work for that separation is already done but not released. Benchmarks appear only as charts measured on 4 SM100 GPUs for prefill and 64 for decode, which is far beyond a small team's budget. The Next Web, citing Bloomberg, notes there is no independent benchmarking yet, so speed and cost claims are DeepSeek's own.

The significance is less about who can run it tomorrow and more about what it reveals. AI labs usually guard serving tricks closely because they decide margins. Publishing them lets hosting companies and large self-hosters copy DeepSeek's methods, which could push the price of running open-weight models down further. The Next Web notes that open-weight models already account for 23% of spending on Vercel's AI Gateway. For builders, the near-term value is learning: the code is a free, real-world textbook on cheap inference, and experts who can explain and apply it will be in demand.

Why it mattersIt hands the wider community the methods behind one of the cheapest AI services in the market, which could lower hosting prices for open-weight models and makes Huawei chips a first-class target for serious open-source serving software.

Business Insider reported on October 19 that Amazon has set up an "Early Rack Receipt" process to take ownership of Nvidia GB300 server racks before they are installed in working data centres. According to internal August emails the outlet reviewed, as summarised by Investing.com and Storyboard18, Amazon expected about $5 billion of racks at first and about $3 billion more each month after that. Some AWS sites have floor space for the machines but not enough utility power, so finished equipment is waiting in intermediary warehouses as what the reporting calls "dark GPUs".

The numbers around it are large. Amazon expects about $220 billion of capital spending in 2026, up from roughly $200 billion forecast in July, according to Coinpaper, and that figure covers all company infrastructure, not only AI. Its trailing 12-month free cash flow fell to about negative $7.6 billion in the second quarter. At the same time, AWS revenue grew 37% in that quarter, its fastest rate in 18 quarters, and its contracted backlog reached about $496 billion. Cleanview Research estimates Amazon has 44 gigawatts of data-centre capacity under development across 16 US states.

Storyboard18, summarising Business Insider, says AWS chief financial officer John Felton joked at an internal meeting in September that the expansion had turned free cash flow negative. Prasad Kalyanaraman, who oversees much of the build-out, reportedly acknowledged that the AI business was producing large cash outflows that could continue for a couple of years, and said AWS struggled to find skilled workers in West Texas and funded nearby housing. An Amazon spokesperson told Business Insider that staging hardware until power is ready is a routine "best business practice" and disputed suggestions that it lacked a plan to bring the sites live.

Several caveats apply. The $5 billion and $3 billion figures come from internal forecasts in emails, not confirmed purchase totals, and we could not access the original Business Insider article directly, so details here rely on syndicated and secondary summaries. The reporting does not say how many racks are idle today or for how long. Amazon's view is that infrastructure costs arrive years before revenue, so negative free cash flow is a timing effect, not weak demand. CEO Andy Jassy has said Amazon expects demand to exceed capacity into 2027.

The bigger lesson is that the AI bottleneck has moved from chips to electricity, transmission lines and skilled construction workers. Even the richest cloud company in the world is waiting on grid connections. That favours anyone who can make AI cheaper to run per watt, anyone who can help data centres and utilities plan, and anyone who can supply skilled electrical and construction labour. For smaller buyers, it also suggests GPU capacity may stay tight and expensive, which rewards efficient models and careful cost control rather than assuming compute will soon be abundant.

Why it mattersIf the largest cloud provider cannot power the GPUs it already owns, compute will likely stay scarce and pricey, and the money in AI shifts toward energy, efficiency and the people who build and wire data centres.