The Lyceum: AI Weekly — Jul 13, 2026

This was the week AI stopped looking like a software race and started looking like an industrial one. The action moved below the model — to who owns the chips, who controls the compute, and who quietly eats the token bill. China demonstrated it can now train frontier-scale models on domestic silicon alone; American companies started switching to Chinese models to save money; and a single blog post about wasted tokens forced everyone to reckon with what agents actually cost. Less demo machine, more plumbing.

The lab that proved you could build world-class AI on a shoestring is now trying to build the hardware to run it — a fundamentally different kind of ambition. (DeepSeek Is Building Its Own Chip — and That Changes the Whole Game)

Reuters reported on July 7, citing three people familiar with the matter, that DeepSeek is developing its own AI chip built for inference — the stage where a trained model generates responses, as opposed to the training that creates it. DeepSeek has spent about a year in talks with chip-design, foundry, and memory partners and has quietly hired chip-design engineers, aiming to cut its dependence on Nvidia and Huawei silicon.

The logic is hard to miss. U.S. export controls bar Chinese firms from buying Nvidia's best chips, and Beijing has pressed its champions to build domestic alternatives. Roughly 70 percent of AI compute demand is now expected to come from inference rather than training, per The Next Web's reporting — precisely where a purpose-built chip earns its keep and where Chinese silicon is already closest to competitive. OpenAI unveiled its own custom inference chip, Jalapeno, with Broadcom last month; Anthropic is reportedly weighing the same move. DeepSeek joining that club carries an extra charge: a lab that rewrote the efficiency playbook for software is now trying to rewrite it for hardware. (DeepSeek Is Building Its Own Chip — and That Changes the Whole Game)

The project is still at the discussion-and-hiring stage — no named foundry, no prototype, no benchmark. The signal to watch is a manufacturing partner. That's when this moves from intention to program, and when U.S. policymakers have to decide whether export controls need to cover chip design, not just the machines that build chips. (DeepSeek Is Building Its Own Chip — and That Changes the Whole Game)

The number that matters isn't the card count. It's what those cards are made of.

According to reporting synthesized by Tech Buzz China, Meituan — better known abroad as a food-delivery giant — disclosed that it trained LongCat-2.0, a 1.6-trillion-parameter model, entirely on roughly 50,000 domestic AI accelerator cards. LongCat-2.0 completed both training and inference on domestic hardware, making it the first publicly disclosed trillion-parameter-class model to do so without Nvidia GPUs. Separately, a 100,000-card Ascend-based supercluster came online this week. (China Turned On a 100,000-Card Domestic AI Supercluster — and the Word "Domestic)

This is the exact outcome U.S. export controls were designed to prevent. The theory was that cutting off advanced chips would cap China's ability to train frontier models. These disclosures suggest China now holds the essential components to train and deploy frontier models on a largely domestic stack. Chinese data centers face fewer constraints on energy use and pay far less for power — paired with large-scale domestic chip availability, that makes the Ascend stack viable for in-country training even as Huawei's chips still lag Nvidia on raw efficiency, per Tom's Hardware. (China Turned On a 100,000-Card Domestic AI Supercluster — and the Word "Domestic)

If similar clusters keep appearing at the provincial level, the story stops being one supercluster and becomes a distributed buildout — the kind lithography controls can't reach. Watch provincial data-center announcements over the next quarter; that's your tell for whether this is a showpiece or a pattern.

The benchmark wars are interesting. The usage numbers are the real story.

On OpenRouter — a platform that routes developer tasks to different models — DeepSeek, Tencent, MiniMax, and Xiaomi are now the four most popular, according to Rest of World. On Vercel, DeepSeek's share of token usage jumped from under 1% to 17% in May, while its share of revenue stayed near 1%. That gap is the tell: Chinese models are winning on price, not just performance. (Chinese AI Models Are Now the Most-Used in the World — and U.S. Companies Are Sw)

And it's climbing the corporate ladder. Per Tech Buzz China's synthesis, Coinbase CEO Brian Armstrong said on X that his company runs GLM-5.2 and Kimi in production, cutting nearly half its AI spend even as usage rose; Microsoft is evaluating a fine-tuned DeepSeek V4 as a cheaper engine for Copilot Cowork; and Uber's roughly 5,000 engineers adopted Claude so fast the 2026 AI budget was gone by April, forcing a hunt for cheaper options. CNBC reported this week that Chinese-built models are gaining traction with U.S. companies as they narrow the performance gap while staying significantly cheaper.

The "China is six-to-nine-months behind" framing no longer holds for agentic coding. The geopolitical risk is real; for many finance teams right now, the cost risk is more immediate. Watch whether Beijing restricts foreign access to its best models — a move that would flip this entire dynamic overnight.

Your AI coding bill is higher than it should be, and now there's a measurable reason why.

Evaluation startup Systima published a detailed comparison finding that Claude Code — Anthropic's AI coding assistant — sends about 33,000 tokens to the model before it reads a single character of your prompt, versus roughly 7,000 for the open-source OpenCode: a 4.7x overhead gap. Tokens are the chunks of text a model processes, and you pay for every one. Systima attributes the extra load to a large system prompt, 27 tool definitions, and heavy scaffolding, and says a real repo setup could push the starting payload to 75,000–85,000 tokens before any work begins. The post, backed by full token logs, climbed to the top of Hacker News. (Claude Code Sends 4.7x More Tokens Than OpenCode Before Your Prompt)

This matters because the overhead tax repeats on every call in a long-running session — an agent grinding through a complex task might make dozens. It gives the enterprise cost problem a concrete mechanism: per The Prompt Factory, Uber capped employee spend on tools like Claude Code and Cursor at $1,500 a month per tool, with its own COO conceding the link between that spend and shipped features "is not there yet." (Claude Code Sends 4.7x More Tokens Than OpenCode Before Your Prompt)

The evidence is one carefully instrumented benchmark from one team — treat the numbers as directional. But if it replicates across other agent harnesses, procurement starts weighing scaffolding as heavily as model quality. Watch for Anthropic's response, and whether rivals turn this into a pricing wedge.

Meta spent years treating AI as a feature stuffed into its apps. This week it signaled a different ambition.

Reuters reported that on Thursday Meta released developer access to its Muse Spark AI model alongside an upgraded version — a move toward being a distribution layer for other developers' work rather than just an app that contains AI. The distinction that matters is product versus production: opening developer access isn't the same as proving the model runs in high-stakes deployments, but it's a real step toward becoming infrastructure. Reuters also reported that Meta discontinued an AI image feature tied to public Instagram accounts days after rolling it out, following concerns about how it worked — a reminder that product design and privacy collide fast.

The strategic read: Meta wants to be the place AI apps plug into, not just the app they live inside. The signal to watch is pricing — if Muse Spark undercuts OpenAI's GPT-5.6 tiers, capable inference commoditizes faster than either lab's margins can absorb. If developer access stays a showcase with no production deployments named, it's a demo, not a platform. (Migrating a Production AI Agent to GPT-5.6: 2.2x Faster, 27% Cheaper)

A food-delivery company casually training a trillion-parameter model on chips America tried to ban; Uber's engineers torching an entire year's AI budget by April and getting a $1,500 allowance like teenagers with a credit limit; and 33,000 tokens of throat-clearing before your coding assistant reads the word "hello." The plot twist nobody scripted: the labs are spending billions to own the chip layer while the actual money leak was the scaffolding all along — and somewhere in Hangzhou, a bookkeeping model is quietly balancing the books that none of this quite does.

Watch the token meter, not the benchmark.

Forward this to the friend who keeps saying China is "years behind" — they've got some reading to do.