Anthropic wants a $2 trillion valuation and roughly $100 billion in proceeds. The investors who sat through its recent briefing are marking the company nearer $1.5 trillion. That spread is the most informative number in AI right now, because it prices the distance between a track record and a set of commitments.

Both figures come from 36kr's 山上 writeup on the IPO. Anthropic had real news for that room. Q3 revenue grew about 45% quarter over quarter, the quarter was adjusted-profitable, and cash flow turned positive for the first time. Q2's growth rate was 147%. The deceleration is what investors reacted to.

Key Numbers

• $518B: minimum ten-year commitments Anthropic has signed with six cloud, compute, and equipment partners. About 80% cannot be canceled.
• $111.1B: the Google portion alone, payable from April 2026 through July 2033 whether the capacity gets used or not.
• 1.7x: revenue Anthropic says it earns per dollar of compute purchased.
• 11%: adjusted operating margin in Q3, excluding stock-based compensation.
• $74B: current annualized revenue, per SemiAnalysis.
• $20.28B: cash, equivalents, and short-term investments on the books as of last December.

The gap the market is pricing

Nearly 20 rounds have put about $125 billion into the company, including a $65 billion Series H in May at a $96.5 billion post-money valuation. Against that, $518 billion of contracted spend is a different order of magnitude. It is the largest single bet any AI lab has placed on its own growth curve.

Commitment	Amount	Window	Cancelable
Google	$111.1B+	Apr 2026 to Jul 2033	No
Amazon	$110.0B	May 2026 to Apr 2036	No
Broadcom equipment leases	~$161.2B	multi-year	Mostly no
Microsoft	$31.4B	Nov 2026 to May 2033	Partly
Others (Fluidstack, Nvidia, SpaceX Colossus 1)	undisclosed	multi-year	n/a

A 1.7x revenue-to-compute ratio paired with an 11% adjusted margin tells you how thin the cushion is. Most of every revenue dollar goes straight back into accelerators. Let the ratio slip to 1.4x and the margin disappears at constant overhead.

The top line is still moving fast. Q1 2026 alone brought in $4.7 billion, more than all of 2025. Q2 hit $11.6 billion with 147% sequential growth, and Q3 landed around $16.8 billion. What changed is the derivative: net new revenue fell from roughly $6.9 billion in Q2 to $5.2 billion in Q3.

Part of the slowdown is self-inflicted. Anthropic tightened enforcement against distillation, and revenue tied to distillation runs 5% to 7% of the total. The rest is competition. OpenAI has been shipping Codex updates into the exact coding market where Claude Code built its lead, and cheap Chinese open-weight models have forced price cuts across the board. SemiAnalysis analyst Joey Brookhart put it plainly to the Financial Times: Anthropic is now in a two-horse race rather than running alone.

The design-manufacture gap

Pat Gelsinger spent 18 years inside Intel, ran VMware, then ran Intel again. He has seen more semiconductor cycles than almost anyone alive, and his read on AI chips during a recent a16z podcast is worth taking seriously.

AI made chip design fast. It did nothing for fabrication. A team with good tooling can finish a new accelerator design in three months. Wafer fabrication takes about three more. Advanced packaging, where the compute die, HBM stacks, and interconnect get assembled, adds months again. Rack and system integration follows. From tape-out to something you can actually deploy at scale, you are looking at nine months at best, and up to 18 months once you count yield ramp, software, and fleet rollout.

The consequence: models now iterate faster than chips can be built.

Picture the trap concretely. A team designs a chip in early 2027 around the assumption that inference means short prompts and fast token generation. By the time the part reaches volume, reasoning models with long context traces dominate the traffic, KV cache eats the memory budget, and the chip's best feature is aimed at a workload nobody runs anymore.

That mismatch is the argument for general-purpose silicon over aggressive specialization, and it is the reason independent GPU vendors still have a lane.

HBM is bad, and copper is running out

Gelsinger's sharpest line: HBM is a bad memory technology, and it is also currently the best one available. Density is low, power and thermals are ugly, and pin bandwidth is boxed in by the package. As accelerator compute grows, more of it idles waiting on data. The industry calls this the memory wall.

Here is the part that matters for anyone planning a fleet. Storage had terrible business economics for two decades. Cyclical pricing meant one good year followed by several bad ones, which made it impossible to fund long-horizon R&D on new materials. AI changed the capital environment fast: memory makers added roughly $2.5 trillion in market value over four years. That money is now chasing ferroelectric materials, capacitor-free cell designs, and tighter compute-memory integration.

Gelsinger is skeptical of one popular answer. Compute-in-memory (PIM) needs architecture changes tuned to specific tasks, which is exactly the wrong property when workloads shift every six months. Putting high-bandwidth memory physically closer to a general-purpose compute unit looks more durable.

Interconnect is the other wall. Copper works fine over short distances. Stretch a high-speed link to 5 meters and the cable, signal processing, and compensation circuitry can cost more than running fiber 100 meters. Optics will land first in scale-up GPU-to-GPU I/O, not in the compute-to-memory path, because every electrical-to-optical conversion burns power and money. Watch 2028 to 2029 for NPO and CPO at volume. Nvidia's NVL72, 72 GPUs wired into one system, is a manufacturing nightmare that exists partly because the optical supply chain wasn't ready earlier.

China's compute map: eight hubs, one invoice

China's answer to the energy-versus-demand mismatch is geographic. Eight national compute hubs and ten data center clusters were designated starting in 2022, and by July 2026 the country reported 2.45 million PFLOPS of intelligent compute (FP16), with the eight hubs and three compute-power coordination zones accounting for over 85% of it. Roughly 1.45 million PFLOPS sits under the national monitoring and scheduling platform, per 铑科技's breakdown.

Hub	Anchor cluster	What it's actually selling
Beijing-Tianjin-Hebei	Zhangjiakou	proximity to research talent and enterprise demand
Yangtze River Delta	Wuhu, integration demo zone	manufacturing, finance, and automotive workloads
Greater Bay Area	Shaoguan	hardware supply chain and terminal manufacturing
Chengdu-Chongqing	Chengdu, Chongqing	western industrial base, new interconnect node
Inner Mongolia	Hohhot, Ulanqab	wind and solar, zero-carbon cluster economics
Ningxia	Zhongwei	large-scale, latency-tolerant training and batch jobs
Guizhou	Gui'an	existing big-data and cloud base
Gansu	Qingyang	new energy paired with very large training clusters

The practical read: if your job tolerates latency, power price per kilowatt-hour decides your invoice far more than FLOPS count does. If your job is interactive, you pay for proximity and you accept the grid you get. The hubs are not interchangeable, and the 85% concentration number means the national grid of compute is really eight regional power-and-cooling businesses wearing data center costumes.

Quick Take: The buildout has shifted from a demand story to a financing story, and three things now decide who survives it: the size of your non-cancelable commitments, whether you own the distribution channel, and whether your org can keep the people who make the models better.

Supernodes: how big, and who signs

Every domestic GPU vendor showed up at China's industrial expo with the same idea: wire more chips into one box.

Iluvatar CoreX packed 144 GPUs and 18 CPUs into a single 44U rack, 10.1 TB of total memory, 100-nanosecond single-hop latency, and 15% better effective bandwidth on MoE all-to-all traffic. Moore Threads, Biren, and Alibaba Cloud built around 128 cards. MetaX shipped 64-card and 16-card variants. Enflame built 64 cards with ZTE handling rack integration.

Configuration	Vendor example	Why that number
144 GPUs	Iluvatar TianGai 300 IXL144	128 for compute plus 16 for redundancy, sized for 10T-parameter models
128 GPUs	Moore Threads, Biren, Alibaba Cloud, ZTE OEX	falls on the power-of-two parallelism boundary most training jobs use
64 GPUs	MetaX Xijing S600, Enflame ES64	matches current customer demand without oversizing the rack
16 GPUs	MetaX S600 Lite	compact cluster for buyers who don't want a hall-sized commitment

The 144-card design says something honest about the state of things. Training parallelizes along powers of two, so 128 cards is the natural compute unit. The extra 16 exist so a single failed node doesn't kill a running job. You are buying redundancy, not throughput.

Then there's the buyer problem. IDC data cited by Reuters puts China's H1 2026 data center AI accelerator shipments at 2.008 million units: Nvidia at 678,000 (34%), Huawei at 572,000 (28%), Alibaba's T-Head around 7%, and everyone else splitting the remainder.

Note the split inside those numbers. GPU shipments dropped 26% year over year to 991,000 units while non-GPU accelerators grew nearly 80% to 1.016 million. Domain-specific architectures are taking real share, and buyers are hedging architecture bets rather than doubling down on one.

I've watched a regional cluster procurement drag on for four months because the buyer wanted next-generation parts inside a budget cycle that closed before the parts existed. The numbers in that table explain why the vendor can wait and the buyer can't.

Distribution beats models

The clearest price war in AI right now is video generation, and it isn't being fought on model quality.

ByteDance merged workflows from Jimeng and Xiaoyunque into Jianying (CapCut internationally), which already had more than 800 million monthly users globally by the end of 2024. Independent labs like Shengshu, Aishi, and Yanyu raised at $2 billion-plus valuations on the strength of exactly the workflow integration ByteDance is now doing in-house. LibLib's LibTV and Flova build project folders, shared asset libraries, and multi-session context so a single generation becomes a full production. Jianying Hub now does the same thing, with distribution attached.

Product	List price (RMB per second, 720p)	Promotional price	Distribution
Vidu Q4 Preview	~0.59	0.09 at launch	API, Alibaba Cloud Bailian
PixVerse V6	~0.60 (no audio)	n/a	Meta catalog video ads, global ad networks
Seedance 2.5	~1.51	0.19 via Jimeng membership tiers	Douyin, Jianying, Xingtu
Jianying AI Ultra	bundled	99 RMB first month, 22,100 credits	full ByteDance funnel

That table is the entire strategic picture. Seedance lists at more than double Vidu and PixVerse, and it still wins on price in practice, because ByteDance can discount through membership credits and recover the money downstream in subscriptions, ad placements, and distribution. A pure model vendor has API pricing and nothing else.

The reason distribution decides this market is the hit rate. Douyin added 221,900 AI short dramas in H1 2026, and 1,055 of them crossed 100 million views. That's a 0.48% hit rate. When almost nothing works, the platform that controls the recommendation feed controls who gets paid.

Common Pitfalls

Treating non-cancelable compute as a variable cost. Anthropic owes Google and Amazon whether it uses the capacity or not. If you model an AI budget with compute as a percentage of revenue, you'll forecast a business that doesn't exist. Put the minimum commitment in the fixed column and build the rest on top.

Sizing your interconnect for the model you plan to train, not the one you can keep busy. A 128-card supernode sitting at 20% utilization costs you power, scheduling complexity, and software adaptation for the extra 96 cards. MetaX shipped a 16-card variant for a reason.

Assuming open-weight models reduce your infrastructure bill. A cheaper competitor cuts your price per token. Your leases don't reprice with it. This is the squeeze hitting Anthropic from Chinese open-weight releases, and the same math applies to any team renting capacity on multi-year terms.

Budgeting AI tooling on launch pricing. Vidu Q4 previewed at 0.09 RMB per second and settles near 0.59 on list. Launch pricing is a customer acquisition instrument, not a cost basis.

Adopting a new accelerator without costing the kernel work. If a chip runs your model but needs new operators and framework adaptation, engineering hours can exceed the hardware discount in the first year. Ask for the porting estimate in writing before the PO.

The org ceiling: iFlytek and Kimi

Two Chinese AI stories this month share a root cause.

iFlytek spent three years as the only major model trained entirely on domestic silicon. The constraint was real: Ascend 910B memory and bandwidth left long-context and code training roughly 5x behind an H200. The 950 part closes that to a claimed 65% to 85% of H200's theoretical efficiency. The problem is timing. Domestic-silicon training stopped being a differentiator the moment Meituan shipped a trillion-parameter model on 50,000 to 60,000 domestic cards on the same day DeepSeek V4 launched.

The financials show the cost of being early and quiet. H1 2026 revenue was 11.62 billion RMB, up 6.52%, with a net loss of 204 million RMB and a non-GAAP loss of 637 million. R&D spending hit 3.007 billion, over 25% of revenue. Government revenue fell 2.65% while business and consumer combined grew 10% to 76% of the total. The company covers 36,000 schools and 77,000 clinics, and its project-based structure was built for those customers. Meanwhile, the office-agent traffic split in June went to Tencent WorkBuddy at 20.97 million visits, ByteDance TRAE IDE at 12.79 million, and Alibaba QoderWork at 7.88 million. iFlytek's Tingjian is an excellent meeting tool that stops at the meeting, and the workflow continues somewhere else.

The number to watch there isn't the model benchmark. It's 3.219 million developers on the API and MaaS platform, up 111% year over year. Platform pull is the only health metric that leads revenue rather than following it.

Then there's Kimi. A 17-year-old named Chen Guangyu joined as a 16-year-old intern, became co-first author on Attention Residuals, and helped build Block AttnRes, which cuts memory complexity from per-layer to per-block. Trained on 1.4 trillion tokens, Block AttnRes matched baseline quality at about 80% of the compute with under 2% inference latency overhead, and added 7.5 points on GPQA-Diamond, 3.6 on Minerva Math, and 3.1 on HumanEval. The technique is inside Kimi K3, a 2.8-trillion-parameter open-weight model.

He left for the Bay Area this month. Discovery isn't the bottleneck for Chinese AI companies anymore. Retention is. When the best people can find you, the constraint moves to what you can offer them on month thirteen.

Sources and further reading

• 36kr / 山上: Anthropic 值2万亿美元吗？
• 36kr / 铑科技: 中国八大算力中心，都握着什么王牌？
• 36kr / 硅基观察Pro: 前英特尔CEO最新判断：绝大部分芯片设计公司将会消失
• 36kr / IT时报: 国产GPU竞逐超节点，谁来买单？
• 36kr / 数智奔流: 夹在字节缝隙中的AI融资明星们
• 36kr / AI前线: 17岁Kimi核心少年奔赴硅谷
• 36kr / 银杏科技: 苦熬三年的刘庆峰，"孤勇者"等待翻盘

One Thing to Remember

Every number in this piece is downstream of one asymmetry: compute commitments are fixed for three to ten years and model demand re-prices every quarter. Whoever holds the shorter-duration side of that trade wins, which is why ByteDance discounts through distribution, why MetaX sells a 16-card rack, and why Anthropic's revenue growth rate matters more to its valuation than its revenue level.

The Bottom Line

If you're raising or deploying capital in AI infrastructure this cycle, model the minimum commitment as fixed cost and assume your compute-to-revenue ratio decays, not improves. Anthropic's 1.7x at 11% margin is a ceiling, not a baseline.

If you're buying accelerators, skip the largest available interconnect scale and size for the model generation you can actually keep busy. A 64-card or 16-card configuration you run at high utilization beats a 144-card rack you run at a fifth capacity while eating the software porting bill.

One thing to watch: the optical interconnect window. Gelsinger points at 2028 to 2029 for NPO and CPO at volume, and the first place it lands is GPU-to-GPU scale-up links, not memory access. If you're planning a multi-year cluster lease that starts in 2027, price in a mid-life interconnect refresh or accept that your copper-based fabric will look expensive against the next cohort.