Nvidia Supply Chain Bottlenecks and Order Cut Rumors
Full research prompt
Investigate the status of supply bottlenecks, including HBM (SK Hynix, Samsung, Micron), TSMC CoWoS advanced packaging capacity, power, and datacenter readiness. Verify the reported order cuts in Taiwan's supply chain: identify the original sources, which companies were named, whether Nvidia or the ODMs (Foxconn, Quanta, Wistron) confirmed or denied them, and the credibility of those reports. Summarize what the evidence implies for near-term shipments.
From Nvidia Earnings Q3 FY2027: Data Center Revenue, Guidance and the AI Cycle
Near-term AI system shipments are still gated by memory, packaging, and power—not by confirmed order cancellations in Taiwan. The mid-November “Nvidia cut” wave was mostly anonymous sourcing and brokerage stress tests after Jensen Huang’s resignation, and it sits badly next to record ODM sales, a raised Nvidia outlook, and a jump in multiyear supply commitments.
Order-cut reports: sources, names, and what was actually confirmed
The actionable rumor started in Taiwan’s Commercial Times (工商時報) on November 10, 2026. Supply-chain sources said Nvidia wanted to slow chip orders, with no size and no date, in the context of scrutiny of Huang and China-related export friction (including roughly 600,000 H200 chips reportedly sitting in Hong Kong pending Beijing approval). The same reporting named TSMC and ASE as the first places a slowdown would show up, and said freed CoWoS could help AMD and ASIC designers (Broadcom, MediaTek, Alchip, GUC). A second Commercial Times piece the same day, plus Economic Daily News, quoted industry people saying orders, capacity, and shipments of chassis, racks, and cooling were unchanged. TrendForce’s write-up of those articles is the cleanest English summary. [1] [2] [3]
A second, louder layer was not a cut report at all. After Huang resigned on November 12, local brokers (covered by Business Today and others) ran hypothetical 10/20/30% order-cut scenarios and scored Taiwan names by “Nvidia exposure.” Companies repeatedly named in that exercise and in related selloff coverage included Wistron, Foxconn (Hon Hai), TSMC, Unimicron, KYEC (King Yuan), Quanta, and substrate/CCL names. One scenario put a 30% cut at a 6.6% hit to 2027 listed-company profits. Those figures are model outputs, not purchase-order changes. [4] [5] [6]
A third thread, from Economic Daily News and unnamed industry sources around November 15–16, was a Vera Rubin volume-timing slip: HBM4 tightness might push the mass ramp from late 2026 into the first half of 2027, with GB300 still the volume platform. That hit CCL, substrate, 800 V power, and optical names in the tape. Substrate suppliers pushed back that Vera and Rubin categories had already entered production on schedule (June and October). Other checks conflict: one supply-chain note blamed motherboard warpage and said the HBM issue was resolved in October, with volume into 2027 and more Blackwell orders as a bridge. [7] [8] [9] [10] [11]
What Nvidia and the ODMs said
- Nvidia did not confirm cuts. On the November 17 call, interim CEO Colette Kress called it “business as usual,” said customers had spent the prior week asking for more shipments, and said the forecast already factors in datacenter readiness, demand durability, and customer access to capital. Q3 revenue was $115.84 billion; Q4 guidance was $130 billion ±3%, above the roughly $125 billion consensus, with zero China datacenter compute assumed. She raised the fiscal 2028 growth outlook to about 75% from 70%. Supply and capacity commitments rose from $279 billion to $390 billion, “primarily related to the multi-year procurement of memory supply.” Commercial Times, citing Kress, said memory and advanced packaging capacity had been secured, and that TSMC executives visited Nvidia that week to close the supply gap. The same day Nvidia announced expanded multiyear deals with Foxconn, TSMC, SPIL, Kinsus, Amkor, and Samsung Electro-Mechanics. [12] [13] [14] [15] [16] [17]
- Foxconn did not report cuts. On November 12 it beat profit expectations, kept a “strong” 2026 revenue view, and guided high-double-digit growth in 2027 AI rack shipments. Chairman Young Liu’s constraint was the opposite of a demand cut: chips, memory, and key components are tight, new capacity will ease bottlenecks only gradually, and demand should still exceed supply next year. On November 17 he publicly backed the next-generation GPU and networking ramp. [18] [17]
- Quanta did not report cuts. It expects AI server revenue up in Q4 even if rack counts dip during the GB300-to-Rubin changeover, because the new systems are higher-priced. Small-volume Vera Rubin racks start this quarter; larger volumes in early 2027; server business up double digits next year. It said it had seen no customer order adjustments tied to worries about US datacenter investment. [19] [20]
- Wistron has not, in the sources reviewed, issued a formal denial. October revenue was up 117.6% year over year, and management expects stronger AI deliveries in November and December. [19] [20]
Credibility. Low for an actual near-term cancellation wave. The only primary “cut” item is anonymous, unsized, and contradicted the same day inside the Taiwan press. Broker 10/20/30% cases are scenario math after a CEO exit. The Rubin timing noise is more plausible as a mix shift and a one-to-two-quarter volume ramp, not as cancelled Blackwell or ODM orders. Nvidia’s own numbers—higher guide, $111 billion increase in supply commitments in one quarter, new packaging and substrate pre-buys—are hard to square with a company winding down Taiwan orders. Reuters’ supplier interviews are about a future risk: a less aggressive CEO might pre-book capacity less hard, which would slow expansions, not cancel work already in the line. [21] [3]
What this means if you are trying to get allocation: do not trade the rumor as if CoWoS slots are about to free up for AMD or custom ASICs. The documented behavior is Nvidia paying to lock the next bottleneck, not releasing it.
HBM is allocated years ahead, and it is starving ordinary DRAM
The three volume suppliers have already sold the next product cycle. On October 29, Samsung told investors its 2027 HBM volume is sold out, with contracts closed at target prices, and that HBM4 sales more than tripled quarter on quarter. On October 28, SK hynix said long-term agreements already cover more than half of 2027–28 capacity, that customer HBM requests are more than double available supply, and that DRAM and HBM are effectively spoken for through 2029 (that line includes volumes still in negotiation, so it is a company characterization, not fully booked revenue). It also said shortages can persist through 2030. Micron has said calendar-2026 HBM was sold out and that agreements cover the vast majority of calendar-2027 HBM bits at prices significantly higher than 2026; more than 75% of its fiscal 2027 shipments were already committed in earlier commentary. [22] [23] [24] [25] [26]
The mechanism is wafer math, not a temporary yield miss. Micron has said HBM3E consumes roughly three times the wafer area of DDR5 for the same capacity, and the gap widens on later generations. TrendForce expects HBM to take about 30% of DRAM wafer starts at the top three by the end of 2027 while producing only about 13% of bits—so every AI stack removes a disproportionate amount of server and PC DRAM. That is why a 16-gigabit DDR5 part was reported moving from about $8.70 to $48 in eleven months, and why a US judge in mid-November let a class action proceed alleging the three suppliers shifted capacity from conventional DRAM into HBM. SK hynix has also started full-scale HBM4 shipments; Samsung has said HBM4E mass production begins in the first half of 2027. [27] [28] [29] [30]
Nvidia’s response is contractual, not technical. The $111 billion quarter-on-quarter rise in supply commitments is mostly multiyear memory. There is also reporting that Nvidia has asked the three suppliers to build base dies for an “NVHBM” on TSMC 3 nm, and separate reporting that it asked TSMC to study its own memory path. Neither changes 2026–27 bit supply. A reported spec trim—Vera Rubin HBM4 bandwidth from 22 TB/s to 19.2 TB/s on Nvidia’s own pages, with no announcement—is consistent with shipping what the memory stack can actually deliver, not with a cancelled platform. [14] [31] [32]
For anyone outside the locked contracts: 2027 HBM is an allocation market. New accelerator programs that were not in those LTAs are waiting on 2028 capacity, paying up, or designing around less memory bandwidth.
CoWoS and substrates are still the factory constraint; new lines do not help this year
TSMC has been explicit that backend supply has not caught demand. On the October 15 earnings call, chairman C.C. Wei said AI demand still exceeds what TSMC can supply, and commentary around that call described a backend shortage that is “still quite big.” Advanced packaging demand has also spread from GPUs into server CPUs and custom ASICs, which is why Commercial Times reported foundries and OSATs weighing another 10–20% CoWoS price increase, with packaging-tool lead times stretching toward six to eight months. TSMC does not publish an official CoWoS wafer count. Industry estimates cluster around 120,000–140,000 wafers a month at TSMC in 2026 (some local-press summaries cite about 130,000), plus 50,000–60,000 from OSAT partners, with a reported path toward roughly double by end-2028. TrendForce has described the gap narrowing from about 20% in mid-2026 toward about 10% by year-end, while other estimates still show a mid-teens shortfall into 2027 because packages are getting larger (5.5-reticle now, 9x-class next). Lead times in trade reports are still quoted in the 52–78 week range, and most 2026–27 slots are described as pre-booked. Treat the wafer counts as estimates, not TSMC disclosures; the direction—demand ahead of supply through 2027—is what management has said. [33] [34] [35] [36] [37] [38] [39]
The capacity Nvidia and TSMC are adding is real and mostly too late for the next few quarters:
- GlobalFoundries will make silicon interposers in Malta, New York, under a five-year, $2 billion agreement. Volume ramp is targeted for the first half of 2028. [40] [41]
- TSMC’s seven-year outsourcing deal with affiliate Xintec covers the on-substrate step of CoWoS inside a Taichung cleanroom. Mass production is targeted for 2029. [42]
- Nvidia is funding SPIL (ASE) with several billion dollars for bumping, interposers, assembly, and test, with dedicated capacity aimed at mid-2029. Amkor agreements in Arizona and South Korea total nearly $3 billion, with one earlier target of more than 300% interposer-wafer capacity by end-2028. [17] [43] [44]
- Kinsus will take Nvidia-dedicated ABF substrate capacity to more than 3.5 times current levels. Production starts in the second half of 2027; the step-up is targeted for the third quarter of 2028. Nvidia is putting more than $1 billion into that expansion. An October package already targeted a 2.5x substrate increase over three years across Ibiden, Samsung Electro-Mechanics, and Unimicron. [45] [46] [47]
Nvidia said Vera Rubin NVL72 and Vera CPUs began volume shipments in the third quarter, on TSMC 5 nm and 3 nm plus CoWoS. That is consistent with Quanta’s “small volume now, larger volume in early 2027,” and with Kress saying hyperscalers brought Rubin into production in about a month. It is not consistent with a cancelled ramp. It is consistent with a ramp that is supply-limited and still mostly ahead of the ODMs’ high-volume quarter. [17] [44]
For a competitor: CoWoS-class slots through 2027 are a negotiation with incumbents who have already prepaid, not an open foundry queue. The practical alternatives are OSAT spillover, a later node on the roadmap, or a package that does not need a full CoWoS-L interposer—and even those are tightening as ASICs crowd in.
Power and shells are now the binding constraint on what actually turns on
Chips can leave Taiwan faster than US (and some European and Asian) sites can energize them. Goldman Sachs, in an October 2026 report shared with Axios, called power “probably the forever bottleneck,” with global datacenter power demand up 170% from 2025 to 2030 and more than 60% of that growth in the United States. It put capital spending across compute, datacenters, and power at about $7.6 trillion from 2026 to 2031. Median time from a US grid-connection application to commercial operation is about five years; the interconnection queue of proposed generation and storage is on the order of 2,600 GW. [48]
Morgan Stanley’s estimate is a 32 GW US datacenter power deficit through 2028, about 34% of projected requirements even after behind-the-meter generation. It argued this concentrates energization among a handful of creditworthy buyers and squeezes smaller developers. Goldman has also said only about 50–60% of planned datacenter capacity comes online on the original schedule over a two-year window. Equipment is a second queue: standard power transformers are quoted at two to three years, generator step-up transformers above 160 weeks in early 2026, and GE Vernova’s gas-turbine backlog runs into the early 2030s. New transformer plants take three to five years; meaningful US capacity is not expected before 2027. [49] [50] [51] [52]
Announced load and buildable load are different markets. AIxEnergy’s week-of-November-15 snapshot showed 175.4 GW of large-load requests across nine US grid operators, of which its model expects only 43.85 GW (25%) to be built. Energy Aspects found that of 129.6 GW of planned US datacenter capacity it tracks, 88.3 GW (68%) had no ground activity, with a median forecast of 11.8 GW reaching full capacity in the next twelve months. Latitude Media separately tracks about 120 GW of signed load versus 22 GW operating in the same footprint—and 537 GW of merely reported load. ERCOT has paused new large-load interconnections while it audits a queue measured in the hundreds of gigawatts. [53] [54] [55]
Nvidia is already pricing this in. Kress said the outlook considers datacenter readiness, and Bloomberg reported she will prioritize customers that have land, energy, and buildings, with hyperscalers likely to drive more sales than neoclouds. Guarantees of up to $118 billion, including about $105 billion tied to an OpenAI-leased Ohio campus, are structured to kick in when shells and power exist—not as a blank check for chip orders that cannot be plugged in. [13] [56] [57]
For a new entrant, a chip allocation without a substation date is not a deployment plan. The buyers who can prepay memory, reserve CoWoS, and show an executed power agreement are the ones who will take the 2026–27 racks.
What the evidence implies for shipments through early 2027
Shipments over the next two to three quarters should keep rising, on a supply ceiling, with a messy platform mix.
- Orders in Taiwan look intact. October sales at the server ODMs were still doubling in several cases (Quanta +104%, Wistron +118%, Foxconn’s best month ever at about $40 billion). Foxconn and Quanta are guiding growth into 2027. Nvidia just raised its own near-term and fiscal-2028 outlook while assuming no China datacenter revenue. [20] [18] [12]
- The volume platform is still Blackwell/GB300. Rubin is in early volume—Nvidia says production shipments started in the quarter, and one call recap put that first production quarter near $20 billion—but ODMs describe the high-volume quarter as early 2027. A slip of Rubin mix into the first half of 2027 would defer 800 V, M9 laminate, and some optical revenue; it would not empty the GB300 line. That distinction is why server assemblers and next-gen component vendors sold off differently. [19] [11] [58]
- The ceiling is components, then power. HBM for 2027 is contracted. CoWoS remains short, with relief from GlobalFoundries, Xintec, SPIL, and Kinsus landing in 2028–29. Foxconn’s own line is that demand exceeds supply next year even as new capacity arrives. Once a rack ships, US grid gear and interconnection—not another Taiwan purchase order—decide when it produces tokens. [18] [40] [51]
- The leadership change is a 2027–28 risk, not a Q4 air pocket. Suppliers told Reuters a successor may be less willing to pre-commit capacity the way Huang did. That would show up as slower expansions and less aggressive spec pushes, which Ming-Chi Kuo has argued is exactly what gave Taiwan its order visibility. It is not visible in current shipments, current guidance, or the November 17 capacity announcements. [21] [17]
Confidence. High that HBM and CoWoS remain structurally tight through 2027, and that power—not a Taiwan order cut—is the deployment constraint in the US. High that no ODM or Nvidia statement has confirmed order cuts, and that the Commercial Times item was unsized and internally contradicted. Medium on the exact Vera Rubin volume quarter: company and ODM statements support “shipping now, larger in early 2027”; anonymous checks still disagree on how much Q4 GPU volume slipped. Low on any specific percentage capacity shortfall (20%, 10%, 15–20%), because those are analyst reconstructions and TSMC does not disclose CoWoS wafers.
Recent Findings Supplement (November 2026)
Recent developments (primarily November 2026) center on market speculation around Nvidia leadership changes rather than confirmed supply disruptions, alongside ongoing HBM and CoWoS tightness with targeted capacity expansions.[1][2]
Order cuts reports originated mainly from Taiwan's Commercial Times (around Nov 10, 2026), citing unnamed supply-chain sources suggesting Nvidia might slow chip orders amid Jensen Huang's resignation (effective Nov 12, 2026, over personal and national security issues). Analyst reports (e.g., Fubon) modeled hypothetical 10-30% cuts impacting 2027 Taiwan-listed profits by 2.2-6.6%, naming exposure in testing (e.g., King Yuan Electronics), CCL/substrates (e.g., Unimicron, Taiwan Union Technology), and ODMs (Foxconn/Hon Hai ~25-56% Nvidia-related per some estimates; Quanta, Wistron, Wiwynn).[3][4]
No confirmations or denials from Nvidia, Foxconn, Quanta, or Wistron appear in reports; multiple outlets (including checks of Reuters, Bloomberg, etc.) found no verified wind-down requests to TSMC or others. Some coverage explicitly notes the claims as unconfirmed or circulating without primary sourcing, with industry comments indicating orders and capacity remain unchanged so far. ODM October 2026 revenues showed strong growth (Quanta +104%, Wistron +118%, combined server makers up ~65-100% YoY).[2][5]
Credibility is low for actual cuts—the reports are speculative and tied to leadership uncertainty rather than demand weakness. Taiwan suppliers and executives have stated that 2027 orders are largely locked in, with complex qualifications making rapid switches difficult.[6][7]
This implies limited near-term shipment impact from "cuts," though short-term volatility and potential shifts toward ASICs (benefiting AMD/Broadcom) could occur; actual data shows resilience.[8]
HBM remains a core bottleneck, with 2027 supply sold out or largely contracted across SK Hynix, Samsung, and Micron. SK hynix reported DRAM/HBM effectively booked through 2029 (with long-term agreements covering >50% of 2027-2028 capacity). Samsung stated 2027 HBM supply sold out at target pricing, with tightness into 2028. Micron noted >75% of fiscal 2027 shipments committed and tighter conditions ahead.[9][10]
Nvidia is reportedly pushing for TSMC involvement in base dies or even 3D DRAM alternatives to diversify from Korean suppliers, while Samsung aggressively expands HBM4 (potentially challenging SK hynix's lead). Packaging/TSV yields and base die capacity add constraints beyond wafer production.[11]
Implication: HBM tightness persists into 2027-2028, limiting accelerator ramps and raising costs; new capacity (e.g., SK hynix P&T6 site targeting H1 2027) offers gradual relief but does not resolve near-term shortages.[9]
TSMC CoWoS advanced packaging capacity faces an estimated ~20% current shortfall, with expansions accelerating via partnerships but new output mostly phased for 2027-2029. TSMC targets ~130k wafers/month by end-2026 (potentially doubling to 260k by end-2028); it is exploring 10-20% price hikes amid long lead times. Recent deals include a $2B agreement with GlobalFoundries for silicon interposers (volume ramp H1 2028) and a 7-year outsourcing pact with affiliate Xintec for on-substrate CoWoS lines in Taichung (production targeted 2029).[12][13]
Nvidia announced multibillion-dollar support (Nov 17, 2026) for partners like Kinsus (substrates >3.5x expansion, production H2 2027+) and SPIL/ASE (advanced packaging lines through mid-2029), plus broader testing capacity doubling between 2026-2027. Vera Rubin systems use TSMC 3nm/5nm and CoWoS.[14][15]
Implication: CoWoS remains the tighter near-term constraint than logic wafers for AI accelerators; outsourcing and funding accelerate relief but new lines lag demand, sustaining allocation pressures and potential prioritization shifts.[16]
Power and datacenter readiness show secondary constraints, with limited new specifics but references to potential delays in 800V DC power supplies and liquid cooling tied to Vera Rubin timing slips, plus US state-level restrictions possibly pushing some builds to 2028. Nvidia's supplier expansions include server assembly support (e.g., Foxconn network).[17]
Overall for near-term shipments (late 2026 into 2027): HBM and CoWoS bottlenecks continue to constrain volumes, with Vera Rubin ramps likely slipping from late 2026 into H1 2027. Speculative order-cut reports have not materialized into confirmed changes—Taiwan ODM and foundry revenues remain robust, and expansions signal sustained demand. However, leadership transition adds uncertainty that could favor ASIC diversification or slower capacity commitments. Evidence points to persistent tightness rather than abrupt cuts, with supply-chain resilience in the short term.[18]
Sources are primarily from November 2026 reporting (Commercial Times, Reuters, company announcements, TrendForce summaries, earnings call coverage); earlier 2026 context (e.g., capacity estimates) is referenced only where it informs recent updates. No major new regulatory shifts beyond ongoing export alignment noted in this period.