2026.09.23 | AC | Cloud | Apsara Day One: Watch the Chip Clock

摘要

1. Whether the 20GW operating-capacity target signals added capex or just relabels a prior goal；
2. What V900's mass-production timeline moving up two quarters actually reveals；
3. Why the bottleneck shifts from chips to HBM once super-nodes scale to a thousand cards；
4. How the 20GW increment splits between in-house chips, HBM, and data-center buildout, and who benefits；
5. Which signals should be tracked next to confirm or falsify the earlier-timeline thesis.

> The 20GW target is simply last year's "tenfold energy consumption" goal restated in absolute terms; the investment envelope has not grown. V900 specs were already out in May — what Day One added was a mass-production date pulled forward by two quarters and a supernode scaled to 1,000 cards. The earlier the timeline, the heavier the reliance on HBM of undisclosed origin. The incremental spend implied by 20GW therefore accrues to HBM and data-center facilities, not to switching and NICs, which Alibaba already designs in-house.

Alibaba's headline number on Day One of the 2026 Apsara Conference was 20GW. CEO Eddie Wu said the global data-center capacity operated by Alibaba Cloud will exceed 20GW by 2032. In the same passage, he conceded that the industry's medium- to long-term demand far outstrips what Alibaba can supply, and that a global shortage across the AI data-center supply chain is capping the pace of compute growth. A bigger target on one hand; an admission that the hardware simply cannot be bought on the other. Read together, the two statements call for a careful audit of what Day One actually delivered: which items are old targets in new clothing, which variables are appearing for the first time, and which binding constraint those new variables point to.

1. 20GW Is an Old Target in New Units — No Step-Up in Investment

20GW points to the same destination as the two previous formulations. At the September 2025 Apsara Conference, Wu said Alibaba Cloud's global data-center energy consumption would grow tenfold by 2032 versus 2022; on the May 2026 earnings call, the phrasing was that the data-center capacity to be built would grow more than tenfold from 2022. This year the metric has shifted to "operated capacity", but the end point is still 2032. What is new about 20GW is that the target is expressed in power capacity for the first time; the investment commitment itself has not been raised.

The funding envelope hasn't moved either — yet the money is going out the door ever faster. The RMB380bn three-year plan was unveiled in February 2025. On the May 2026 earnings call, management said AI infrastructure investment would far exceed RMB380bn, and in August it disclosed cumulative spending of roughly RMB190bn through end-June. The Day One keynote offered no new CapEx figure. According to Alibaba's results announcements, FY2026 CapEx came to RMB126.06bn, with a free cash outflow of RMB46.61bn. In the June 2026 quarter alone, CapEx hit RMB67.68bn, up 75% YoY, and free cash outflow reached RMB44.67bn — a single quarter's shortfall nearly matching that of the entire prior fiscal year. The takeaway: spending is running ahead of official guidance. If capital were truly the binding constraint, the obvious move on Day One would have been to lift the investment number. Alibaba did not.

The constraint Alibaba itself keeps citing sits on the supply side. On the November 2025 earnings call, management said server deployment could not keep pace with the growth in customer orders. In May 2026, Wu said in-house chips were still deployed at limited scale on Alibaba Cloud, mainly due to capacity constraints. In August, he added that contracted orders suggest compute demand will continue to exceed supply. Our understanding is that both Alibaba Cloud and Volcano Engine have largely stopped leasing compute to outside customers, which corroborates the "demand exceeds supply" narrative. The funding side tells the opposite story: Alibaba's August share placement raised net proceeds of about HK$79.7bn. The cash is already in hand; what is missing is compute that can actually be procured and racked. Whether Day One beat expectations therefore hinges on whether it addressed the supply constraint — and the answer lies on the chip side.

It also matters whose balance sheet the 20GW sits on. Wu specifically referred to data-center capacity "operated by Alibaba Cloud", and said the company would invest alongside all of its partners. On the May earnings call, he noted that Alibaba may in future co-build data centers with data-center service providers by selling them T-Head AI servers. Taken together, we believe part of the 20GW will be financed by partners and operated by Alibaba Cloud. The distinction is critical for modeling: backing out Alibaba's own CapEx from 20GW would overstate it. Partners' balance sheets absorb part of the build, while Alibaba gains operated capacity and the right to use the compute.

2. The Real News Is the Timeline: V900 Volume Production Pulled Forward by Two Quarters

V900 specs were public in May; what Day One changed was the timing. At the Alibaba Cloud Summit on May 20, 2026, T-Head unveiled its Zhenwu chip roadmap for the first time, slating V900 for 3Q27 with 3x the performance of M890, 216GB of memory and 1,200GB/s of chip-to-chip interconnect bandwidth. The specs announced on Day One are identical to May's. The change is that V900 will now enter volume production and go on sale in 1Q27, with the next-generation Panjiu supernode server built on V900 launching in the same quarter — two quarters ahead of the May roadmap. For investors, the timeline carries more signal than the specs: specs define what a single card can do, while the timeline dictates when capacity, packaging and memory allocations must be locked in with hard cash.

The supernode has jumped an order of magnitude in scale, and in-house silicon fills in the final pieces. Launched in May, M890 paired with ICN Switch 1.0 supports full-bandwidth interconnect across 64 cards. V900, unveiled on Day One, extends full-bandwidth, high-speed interconnect to 1,000 cards via ICN Switch, supports unified memory addressing, and allows a single cluster to scale to 500,000 cards. The new server integrates four categories of in-house chips: the Zhenwu V900, ICN Switch, the Panmai SmartNIC and the Zhenyue SSD controller. T-Head also disclosed its server CPU roadmap for the first time: Yitian 720 and Yitian 730 will launch in 3Q27, with Yitian 730 the first to adopt a fully in-house CPU microarchitecture, delivering up to 1.4x the single-core performance of Yitian 710. Further out, Yitian 750 will connect directly to Zhenwu chips over the ICN bus. We understand that Alibaba's proprietary networking protocol has been running in Alibaba Cloud's production environment for about a year, with data-communication chips the logical next step — and Day One's announcements sit squarely on that path.

Huawei pulled its timeline forward the same week — no coincidence. At Huawei Connect on September 17, Huawei announced that Ascend 960DT would be ready in 1Q27, three quarters ahead of schedule. The Ascend 960 SuperPoD unveiled alongside it packs 4,096 cards into a single node with 1PB of HBM capacity. Within five days, both companies had moved their next-generation chips into the same quarter. Timelines typically move forward for only two reasons: R&D finished early, or supply-side visibility improved. Either way, the story ultimately runs through manufacturing. Chip design, interconnect protocols and system integration are all things Alibaba can do itself — but there is one input it cannot make, and it did not spell that out on Day One.

3. A Faster Clock Shifts the Bottleneck to HBM

Each Zhenwu generation carries 1.5x the memory of its predecessor. The 810E officially debuted in late January 2026 with 96GB of HBM2e, 700GB/s of chip-to-chip bandwidth and seven ICN ports per chip. The M890, launched in May, raised memory to 144GB, and the V900's 216GB is 2.25x that of the 810E. Across three generations spanning roughly a year, memory capacity has more than doubled — yet Day One disclosed nothing about the generation, stack count, source or packaging of the V900's HBM. Huawei, by contrast, keeps its memory specs in-house: the Ascend 950DT uses its proprietary HiZQ 2.0 (144GB), and according to Huawei Connect disclosures, the Ascend 960DT will carry up to 288GB of on-package HBM. In 1Q27, two domestic chips with more than 200GB of memory per card will reach volume production or readiness — and only one of them has its memory spec defined by the chipmaker itself.

Memory capacity is not a free design choice; it is reverse-engineered from the model roadmap. The roadmap disclosed on Day One shows Qwen4 already in training, with Qwen4.5 and Qwen5 scaling to 5–10 trillion parameters. As a rough estimate, storing the weights of a 10-trillion-parameter model in FP8 takes about 10TB. We understand that for mixture-of-experts (MoE) models, all experts must stay resident in memory during inference, with roughly 30% headroom on top, bringing the total to around 13TB. On that basis, serving one such model would require about 136 810E cards, 91 M890 cards or 61 V900 cards. Viewed as a chain, model size sets the floor for per-card memory, and per-card memory determines how much HBM must be locked in for every server racked. At 1,000 cards, a single V900 supernode implies roughly 216TB of HBM — a volume that has to be scheduled more than a year in advance.

On the domestic HBM front, there is still no public volume-production guidance. In the interim report CXMT released in August, the products listed as in volume production were DDR5 and LPDDR5/5X; HBM has yet to enter its commercial lineup. Public reports indicate the company plans to mass-produce 12-high HBM3E in 2027. Scarcity across the AI compute chain has migrated from compute to memory allocation, and V900 brings that problem home to the domestic supply chain. Alibaba's own product design is also working around memory limits: the next-generation CPFS launched on Day One cuts AI storage costs by up to 69%, while the new KVCacheStore is built around "trading storage for compute", offloading inference KV cache to the storage tier. Such designs are themselves evidence of tight memory — if HBM were plentiful, there would be no need to move the cache out.

By extension, how many V900s can be racked in 2027 depends on how much HBM Alibaba can secure. "Volume production and sales in 1Q27" means the product can go on sale; shipment volume is set by HBM allocation, and the pace at which the 20GW fills up in 2027–28 is tethered to this single input. There is also a reverse read: had Alibaba already locked in captive or long-term HBM supply, disclosing it would have been a clear valuation positive — yet it stayed silent on Day One. Only two explanations are plausible: either supply has not been secured, or the supplier prefers not to be named. Both point to the same tracking item. Following the "can it be built in-house" thread downstream, we can now map where the 20GW increment flows.

4. The Increment Flows to HBM and Facilities — In-House Links Stay In-House

Alibaba supplies all four chip categories inside the supernode, squeezing out third parties. ICN Switch, the Panmai SmartNIC, the Zhenyue SSD controller and Yitian CPUs cover switching, networking, storage control and general-purpose compute respectively, and within the V900 supernode all four come from T-Head. Third-party opportunities in Alibaba's incremental spend are therefore concentrated in the clusters where Alibaba continues to buy external GPUs. Nvidia remains the top-tier partner at this year's conference, and heterogeneous clusters will coexist for a long time, but we expect the share of external GPUs in Alibaba's new compute to decline gradually as V900 ramps — and third-party switch and NIC share tied to those clusters to narrow accordingly.

Whether optical modules benefit depends on the cross-rack medium in the 1,000-card supernode. In our previous report, we argued that optical module demand roughly equals card count × supernode penetration × per-card attach ratio, and that only cross-rack supernode penetration drives optical interconnect — whereas M890's 64-card full-bandwidth domain is largely contained within a single rack. V900 expands the full-bandwidth domain to 1,000 cards; given realistic per-rack power and cooling limits, going cross-rack is unavoidable. Alibaba, however, has not disclosed whether cross-rack links will be copper or optical, or pluggable modules versus near-packaged optics (NPO). Huawei offers a useful reference point: the Ascend 960 SuperPoD uses about 5,500 NPO optical engines in place of the roughly 48,000 800G optical modules it would otherwise have required. In other words, cross-rack does not necessarily translate into incremental pluggable modules. Per company disclosures, overseas revenue accounts for 90.58% of the total at Innolight and 96.16% at Eoptolink, so Alibaba's increment carries limited weight for either. How much the domestic optical interconnect supply chain captures will remain unclear until Alibaba discloses its interconnect architecture for the 1,000-card supernode.

Scarce upstream inputs benefit by allocation; non-scarce ones only by order volume. According to Omdia, CXMT ranked as the world's fourth-largest DRAM maker by 4Q25 revenue. If V900 adopts domestic HBM, CXMT is the obvious candidate supplier; the confirming signal would be its HBM appearing among T-Head's volume-production part numbers. Until then, any pricing of domestic HBM rests on assumptions. SMIC's upside hinges on V900's still-undisclosed process node. Domestic 2.5D packaging capacity is not a bottleneck, so what SJ Semiconductor gains from V900 is order volume with limited pricing power. Both sit upstream, but links governed by allocation and links with ample capacity earn profits of an entirely different nature.

Facilities and power distribution benefit on volume, but whose books carry the cost depends on the partnership model. AtHub has long built custom data centers for Alibaba, and in 2018 and 2019 it repeatedly announced receipt of letters of intent from Alibaba for data-center capacity. As we noted in our previous report, AtHub's top five customers account for 79.07% of revenue, while construction in progress is only about 0.5% of fixed assets — squarely in an investment lull. If the co-build portion of the 20GW goes to operators like AtHub, construction in progress will move before revenue does, making it the earliest observable signal. Under a co-build model, partners fund construction and Alibaba Cloud gains operated capacity, so part of the depreciation burden lands on the partners' financial statements rather than Alibaba's.

T-Head's "external revenue" needs to be read in three layers. In our January 23 report, we argued that competitors will not buy chips carrying a hyperscaler's brand, and that the key metric for T-Head's independence is the share of revenue from non-related parties. Based on Day One and earnings-call disclosures, that figure blends at least three layers. The first is the "more than 60% of T-Head chips serve external commercial customers" figure Alibaba disclosed on its March earnings call, which refers to the portion of Alibaba Cloud's public and hybrid cloud offerings consumed by external customers — in essence, still cloud revenue. The second is servers to be sold to co-build data-center partners: external in form, but in substance serving capacity operated by Alibaba Cloud. Only the third layer — chips and servers sold to genuine third parties — is truly external; as of June 2026, the Zhenwu series served more than 650 enterprise customers. Note also that T-Head has been folded into the "AI Cloud and Compute Services" segment since the June quarter, making standalone figures harder to discern. We suspect V900's external sales and Alibaba Cloud's internal use draw on the same HBM allocation: the more T-Head sells outside, the less Alibaba keeps for itself.

For Alibaba itself, Day One supplied a volume anchor but still no annual cadence. As discussed in our earlier report, the pricing anchor for large Hong Kong share placements has shifted to whether the use of proceeds can be verified. Day One set 20GW as the end point but gave no annual gigawatt additions, so verifiability has not materially improved. The fundamentals read well: in the June quarter, AI Cloud and Compute Services revenue reached RMB48.44bn and adjusted EBITA RMB5.63bn, up 133% YoY, with segment margin rising from 9.1% in the prior quarter to 11.6%. The cash-flow gap, meanwhile, is widening. Net the two, and what the market really needs to verify is whether margins can keep expanding once depreciation starts to hit in earnest — which hinges on how many V900s are actually racked, and that brings us back to HBM.

5. What to Track: HBM Source First, Then November CapEx

Five signals, in chronological order. First, whether the breakout sessions on September 23–24 fill in V900's memory type, process node and supernode interconnect architecture. Second, the source of V900's HBM ahead of 1Q27: if CXMT's HBM enters T-Head's volume-production part numbers, Alibaba's HBM constraint will ease materially — and only then will CXMT's upside be realized. Third, the September 2026 quarter results due in November: watch CapEx, whether investment guidance and annual cadence are updated, the direction of segment margin relative to 11.6%, and whether model and application services ARR can progress from RMB16bn in August toward the year-end target of RMB30bn. Fourth, whether construction in progress as a share of fixed assets in AtHub's 3Q and annual reports moves away from roughly 0.5%. Fifth, whether Alibaba Cloud resumes leasing compute to outside customers — a direct signal that supply is easing.

The falsification conditions are equally clear. If Alibaba discloses captive or long-term locked-in HBM supply before 1Q27 and racks V900 instances at scale on schedule, our thesis — "the timeline moves forward, the bottleneck shifts to HBM" — would need to be walked back. If V900 volume production slips, or deployment falls well short of product guidance, the thesis is borne out. Both scenarios will play out within the next two quarters; there is no need to wait until 2032.

Upside risk comes from supply easing faster than expected. Faster-than-planned gains in CXMT's HBM yield and capacity, looser overseas supply of HBM or accelerators to China, or an annual build schedule and higher investment guidance from Alibaba would all convert Day One's pulled-forward timeline into shipments sooner — in which case our ramp estimates would prove conservative.

Downside risk centers on execution timing. A delay in V900 volume production, further HBM price increases driving up per-card costs, slowing demand prompting CapEx cuts, or a larger-than-expected overseas share of the 20GW diluting the benefit to the domestic supply chain would all weaken our conclusions. Disclosure risk also warrants attention: with T-Head folded into the segment, standalone data on in-house chips may shrink further, making tracking harder.

The above content does not constitute any investment advice. Invest at your own risk.