Split capacity risk from price risk in AI buying

On 22 September 2026 two seller-side statements pointed in opposite directions. Alibaba’s CEO said the industry’s mid-to-long-term demand far exceeds his company’s supply capability and that supply-chain shortages are capping compute growth. The same day, Tencent pushed 2K image generation to ¥0.15 per image. Prices are falling while capacity tightens; a procurement plan that watches only one of the two will misprice the problem.

The seller’s CEO says capacity is short

Buyers and analysts talk about compute scarcity constantly. A seller’s CEO admitting his own capacity shortfall is rarer. In his Apsara 2026 keynote, Alibaba Group CEO Wu Yongming said customer AI demand is very strong and that the company is bringing compute online as fast as it can, then added: the industry’s mid-to-long-term demand far exceeds our supply capability, and the current global shortage in AI data-centre supply chains is limiting how fast our compute can grow (Sina Finance keynote transcript, 22 September 2026).

The same speech set a target: more than 20GW of Alibaba Cloud-operated data-centre capacity by 2032. The weight of this information comes from who said it. The binding constraint is no longer only whether the money exists, but whether power, memory and supply chains can arrive on schedule.

Keep the statement and the data apart. A capacity shortfall is management’s read of supply and demand, not a contractual capacity commitment. The 20GW figure is a 2032 target, not built capacity. What a buyer can take from it is a risk register entry about lead times.

The same day, the image cost floor dropped again

That morning, Tencent released the Hunyuan Image 3.5 preview, priced on the Tencent Cloud API at ¥0.15 per 2K image, or $0.024 internationally, billed on output images only (TMTPost, 22 September 2026, 10:53). The same report notes the model had run in a two-month grey release inside Yuanbao, during which image-generation requests rose more than 50%.

Label the comparisons carefully. Tencent’s quality claims, an internal blind test by several hundred professional designers, a 30% gain over the previous version and parity with ByteDance’s Seedream 5.0 pro, all come from a vendor-organised test with no third-party reproduction. They are vendor claims, not an industry baseline. What is verifiable is the pricing and billing shape: per image, output only.

Models set the price, power and supply chains set capacity

Put the two statements together and price and capacity emerge as separate constraint lines.

Falling unit prices come from model commoditisation and inference optimisation: capabilities are converging while caching, routing and quantisation keep cutting the cost of each task. The capacity constraint comes from the physical side: power, memory, chips and data-centre lead times. Different drivers, so they do not move together.

That is why waiting for one more price cut is an incomplete plan. It saves unit cost and can cost you a queue position. Ramp payment data across roughly 70,000 businesses shows average token prices down 41% from the March peak in August, while per-employee AI spend at the heaviest 1% of users fell nearly 10% over the same period, so price and usage signals were not in step (TechCrunch citing Ramp, 9 September 2026).

Limit the inference. That is a single August reading, and the same index showed zero growth in August through October 2025 before recovering at year end, so one month cannot settle a trend. It supports one conclusion only: price and capacity cannot substitute for each other, and cannot be extrapolated along the same curve.

Our judgment: split the procurement list in two

Three moves that survive contact with procurement:

1. Split the contract in two. One column for unit price and billing scope, one for capacity terms: reserved instances, priority, burst ceilings and remedies when capacity is unavailable. A contract that only negotiates unit price does not solve capacity.
2. Model the two separately. Forecast unit price from technology trends and capacity from peak demand. Extrapolating capacity from a price curve merges two risk classes into one line and leaves you unable to attribute the failure.
3. Book generative assets per unit. A price like ¥0.15 per 2K image can go straight into a budget, but verify the billing boundary: reference images, resolution tiers and failed retries. The boundary overspends more often than the unit price.

One action sits outside the contract: log the seller’s own capacity statement in your risk register. It is not grounds for a claim, but it is enough to justify two decisions, reserving volume early and preparing a second supplier.

One question for buyers

If your agent usage doubles next quarter, which limit do you hit first: budget or capacity? The fixes differ. Budget is model tier and caching; capacity is volume reserved in advance. Not knowing which one applies means spending against the wrong bottleneck.

References: Sina Finance, full transcript of Wu Yongming’s Apsara 2026 keynote (22 September 2026, 10:57; the capacity quote, the 20GW target, Zhenwu V900 and the 5-10T parameter plan all come from the transcript), finance.sina.com.cn/hy/hyjz/2026-09-22/doc-inissitf7053669.shtml (curl HTTP 200) | TMTPost, “Tencent Hunyuan Image 3.5 preview released, Tencent Cloud API priced at ¥0.15 per image” (22 September 2026, 10:53; the ¥0.15 and $0.024 prices, output-only billing, the blind test and the grey-release request growth are vendor launch claims from that report), www.tmtpost.com/nictation/8148434.html (curl HTTP 200) | TechCrunch, 9 September 2026, citing Ramp payment data (August average token price down 41% from the March peak, top 1% per-employee AI spend down 10%, roughly 70,000 businesses), techcrunch.com/2026/09/09/ai-spend-per-employee-slumped-at-top-firms-in-august-summer-doldrums-or-a-warning-sign (curl HTTP 200) | Tencent Cloud Hunyuan product page (current product line; rendered client-side, so pricing figures follow the launch report), cloud.tencent.com/product/hunyuan/pricing (curl HTTP 200). Vendor-stated figures and vendor-organised blind tests are not third-party reproduced; recorded here as vendor claims only.