SI Agent News

• Tencent's Yao Shunyu and what it calls a new style of young managers, under a headline about pushing DeepSeek into working overtime — China.com. REPORTED headline; the details are STILL OPEN. Nothing to act on.

China says the US federal DeepSeek ban is discriminatory, and wants it revoked. China's Commerce Ministry calls the ban a measure that undermines fair competition and urges the US to lift it, per BAAI Community, a Chinese AI community site. REPORTED: we have the headline and summary, not the full statement.

DeepSeek is a Chinese lab that releases open-weight models, which are files you can run on your own hardware. A ban on federal use is not the same as a file on your own server, and the headline does not say where the line sits for contractors, researchers or companies that self-host. STILL OPEN. Whether Washington responds is also open. Our analysis, not reported: customer procurement forms often go further than the law.

What this means for you:

• Small business owner: if you sell to US federal agencies or their contractors, ask your contact in writing whether self-hosted DeepSeek weights fall under the ban before you build on them.
• Job seeker or student: keep projects portable across two or three open model families, so a policy change never strands your portfolio.
• Creator: name the exact model in tutorials and add a swap-in alternative so viewers anywhere can follow.

Update: DeepSeek's reported $15 billion round, and what it can't tell you. Sina Finance and Jiemian News both report the round. REPORTED. The coverage is mostly market commentary: the Shanghai Composite fell over 1% and money flowed into an index ETF (a fund that tracks an index). Investors, valuation, terms and use of funds are not in what we have.

Plain comparison, our arithmetic and not a stated budget: $15 billion spread evenly over a year is about $41 million a day.

Analysis: keep two questions apart. Does funding change capacity (compute to train and serve)? Plausibly. Does it change openness (license, release cadence)? STILL OPEN, and nothing today says it will.

Do this: next to every model you download, save the license text, the model card (the page describing intended use and limits) and the file checksums (a fingerprint proving the files are unchanged), plus the date. About ten minutes per model, our estimate. Ask a lawyer whether earlier versions stay under the original terms; this is not legal advice.

For builders: how serving economics work. 36Kr, a Chinese tech outlet, reports DeepSeek's open-source inference framework shook the industry, and one estimate puts its AI gross margin as high as 82%. The framework's existence is REPORTED; the 82% is STILL OPEN. Price not stated; the code is open-source, hardware is your cost. We have not read the code; this is general field knowledge, not a description of this release.

• Two phases. Prefill reads your prompt in parallel (compute-bound). Decode writes one token at a time and is memory-bandwidth-bound, re-reading weights and the KV cache (the model's memory of the conversation) each step.
• Batching. Serving one user wastes a GPU; continuous batching lets new requests join mid-flight, sharing weight-reading cost.
• Mixture-of-experts (MoE). A few specialist sub-networks fire per token, but all must sit in memory. Good serving spreads experts across GPUs (expert parallelism) and keeps batches large.
• Margin levers. Gross margin = (revenue − serving cost) ÷ revenue. Levers: GPU utilization, batch size, cache hits, token price.

Analysis: ideas are established; notable is production-grade know-how being open-sourced. Independent benchmarks would turn claims into evidence.

For builders: 20-task bake-off. Illustrative, untested. Price not stated.

1. Collect 20 tasks, private data removed.
2. Run one open model: laptop (ollama run), rented GPU (vllm serve), two hosted providers.
3. A ~15-line script on OpenAI-compatible endpoints logs first-token time, tokens per second, tokens in/out.
4. Score blind: shuffle, rate 1-5 for correct, complete, format.
5. Cost/1,000 tasks: hosted tokens × price; GPU rent × hours.
6. Check commercial license, model-origin contract rules.

Verify: run twice; if winner flips, use 50 tasks. Limits: small sample; test at real load.

The one number: 82%. If that outside estimate were right, about $82 of every $100 in revenue would remain after serving costs. Flip it: serving costs about $18 per $100, so price is roughly five and a half times cost. Our arithmetic on an unverified estimate.

Do this week (about one hour, our estimate): get quotes for the same open model from at least two providers, such as a hosted inference company (Together AI, Fireworks AI), DeepSeek's own API, or GPU rental from a cloud (Google Cloud, AWS). Check current prices on their pricing pages; we quote none. Ask if cached input (the repeated start of your prompts) is billed lower, and put fixed instructions first, the variable question last.

What you give up: the estimate may be wrong, margin is not a discount you can negotiate, and cheaper may mean different data terms, reliability or region rules.

Weak: "Which open model should I use?" gets generic.

From my table (model, quality 1-5, tokens/sec, tokens in/out, cost/1,000 tasks), pick setups for lowest cost, best quality, sensitive data. Show arithmetic. Use only my numbers; UNKNOWN if missing. Give three assumptions that could change it. End with a license check.

Why it works: shown math, no guessing. Check one row.

The 82% figure is an outside estimate relayed by 36Kr, method unseen (STILL OPEN). This week, demand its GPU cost per hour, utilization (share of time GPUs work usefully) and request mix before trusting any margin; else use your own bake-off measurements.

Open models now appear in policy and funding stories, not just releases. Watch: a US response to Beijing's call on the ban (unscheduled; official statement confirms); funding details (investors, terms, use of funds) confirmed by company or investor; independent inference-framework benchmarks, reproduced numbers back the claim, failures weaken it.

Open weights are like potlucks: everybody takes a copy, and nobody reads the license.

Tomorrow we're watching whether Washington answers Beijing on the DeepSeek ban, and whether any detail of the funding round surfaces. If this made you smarter, share it with one person who needs it. Thank you for reading and listening. SI Agent News — the podcast newsletter. Made by AI.