A GPU server lead time is set in several places at once. NVIDIA reports lead times "of more than 12 months" for certain supply [1], and Super Micro says its manufacturing lead times have "extended beyond twelve months for some products" [5]. A reseller, Viperatech, quotes 8 to 14 weeks for standard HGX B300 builds [32]. Those weeks sit on top of capacity that chip and memory makers reserved well in advance [2, 12].
Filings and earnings calls from November 2025 to November 2026 place the delivery month at six stages: chip supply, advanced packaging, high-bandwidth memory (HBM), substrates and testing, the system builders, and the powered building. No maker in these sources gives a lead time for a named server model or a GB300 NVL72 rack. For the building, the Department of Energy put lead times for critical grid equipment at two or more years in August 2026 [23].
Why GPU lead times are so long
| Stage | What the sources show | As of |
|---|---|---|
| Chips and supply | NVIDIA's supply commitments rose from $119 billion to $279 billion in one quarter, primarily for memory and manufacturing facilities [2] | 26 July 2026 |
| Advanced packaging | TSMC's packaging capacity was "limiting my customers' growth" in July, with "a little bit shortage on the back end" in October [25, 26] | July and October 2026 |
| HBM | Micron has completed agreements for the vast majority of its calendar 2027 HBM bit supply [12] | 30 September 2026 |
| Substrates, testing, optics | New capacity targeted from February 2027 to the end of 2028 [4] | 21 October 2026 |
| System builders | Dell's AI server backlog was $95 billion, and shipment timing varies with customer readiness [8, 9] | 31 July 2026 |
| Powered buildings | 80.4% of capacity under construction in primary North American markets was preleased [28] | First half of 2026 |
Each stage is reserved in advance. NVIDIA says that to secure supply it has "paid premiums, provided deposits, and entered into long-term supply agreements and capacity commitments" [1]. By this article's sum, $175 billion of its commitments fall in its fiscal years 2028 and 2029, which run from early 2027 to early 2029 [2]. The Q3 2026 compute market review covers the same figures as quarterly news.
What makers say about GPU server lead times
NVIDIA's quarterly report for the quarter to 26 July 2026 says it is "currently experiencing certain supply constraints" while it ships both Blackwell and Rubin systems [2]. Super Micro describes the same practice as NVIDIA: component orders placed ahead of historical lead times, with premiums and deposits [5].
NVIDIA's customer advances rose from $160 million to $2.8 billion in the six months to 26 July 2026 [2].
B300 and GB300 NVL72 lead times
No maker in these sources gives one. Viperatech's site says "HGX B200/B300 lead times are 8-14 weeks for Golden SKU" and "Custom BOMs exceed 20 weeks" [32]. That is one seller's claim, viewed 7 November 2026.
CoWoS capacity in 2026
Advanced packaging joins the GPU die and its HBM stacks on an interposer. TSMC's chief executive, C.C. Wei, said on 16 July 2026: "our packaging capacity is so tight that now it's limiting my customers' growth" [25]. On 15 October 2026, describing a back-end deal with GlobalFoundries, he said: "Definitely, we have a little bit shortage on the back end" [26]. He called CoWoS "still the number one choice for our customer" [26].
TSMC gave no CoWoS capacity figure and no size for the shortfall in either call [25, 26]. Published figures are estimates. TrendForce News reported in June 2026, citing Economic Daily News and institutional investors, that TSMC's monthly CoWoS capacity could reach 120,000 to 140,000 wafers in 2026 and that the supply gap could narrow from around 20% to about 10% by year-end [31]. TrendForce expects the shortage of 2.5D packaging capacity "to begin moderating by 2027" [31].
HBM supply in 2027
HBM takes more factory space than ordinary memory. Micron's annual report says "HBM requires a higher number of wafers and more cleanroom space to produce the same number of bits as conventional DRAM" [11]. On 30 September 2026 Micron said it had "completed agreements for the vast majority of our calendar 2027 HBM bit supply" [12]. For memory and storage overall, it expects conditions "to be much tighter in calendar 2027 and 2028 than they were in 2026" [12].
SK hynix and Samsung describe the same conditions. By 28 October 2026 SK hynix had signed long-term agreements with 14 major customers, covering more than 50% of its production capacity for 2027 and 2028 [33]. Samsung expects demand for server DRAM, enterprise SSDs and HBM to keep the market "undersupplied" [14].
Memory also limits the servers built around the GPUs. Super Micro says shortages of memory, storage, GPUs and CPUs "affected the timing of certain of our product deliveries" in fiscal 2026 [5]. HPE says memory constraints "may persist until memory vendors transition greater production allocations towards high performance memory components required by AI workloads" [10].
Substrates, testing and optics have their own dates
NVIDIA's announcement of 21 October 2026 dates new capacity for substrates, interposers, testing and optical engines [4].
| Input | Supplier and location | Target | Date |
|---|---|---|---|
| Testing | Singapore operations | High-volume production | February 2027 |
| Testing | KYEC, United States, with a $1.1 billion capacity reservation advance | High-volume production | April 2027 |
| Substrates | Unimicron, Taiwan | Adds capacity, then reaches full capacity | Second quarter of 2027 to mid-2028 |
| Optical engines | SPIL, Taiwan | 15x capacity | Mid-2027 |
| Substrates | IBIDEN Ono Plant, Japan | Additional capacity begins ramping | First quarter of 2028 |
| Interposers | Amkor, South Korea | More than 300% more interposer wafer capacity | End of 2028 |
Source: [4].
Substrate capacity for NVIDIA's products is to rise 2.5x over three years, and total testing capacity is to double from 2026 to 2027 [4]. These are forward-looking targets. Most begin in 2027, so systems delivered in late 2026 draw on capacity already in place.
System builders: backlog, readiness and acceptance
Dell booked $60.9 billion of AI server orders in the quarter to 31 July 2026, recognized $16.4 billion of AI server revenue and ended with a $95 billion backlog [8]. Super Micro received new orders "in excess of $60 billion" in the quarter to 30 June 2026, "expected to be delivered over future quarters" [6]. Neither company says how its backlog is scheduled.
Dell ties the timing of its AI server shipments to "the varying stages of customer readiness, and the frequency of component part updates or transitions" [9]. HPE reports higher inventory due partly to "longer customer acceptance timelines on AI-related orders" [10]. Delivery is one date. Acceptance is another.
Powered buildings set the last date
NVIDIA's finance chief wrote in August 2026: "Securing land, power and shell for data centers has become the next critical phase in the AI infrastructure buildout" [3]. TSMC now sends its own teams "to check on the data center build-out progress and power supply situation in the U.S." [26].
CBRE tracks eight primary North American markets [28, 29]. In the first half of 2026 vacancy was 1.4%, 80.4% of capacity under construction was preleased, and less than 1,500 MW remained available, about six months of demand [28, 29]. In February 2026 CBRE wrote that sites "offering power access within 18 to 36 months for greenfield development are highly sought after" [27]. The Department of Energy puts lead times for critical grid equipment, including transformers and circuit breakers, at "two or more years" [23].
Inside the building, CBRE reported in August 2026 that ready-for-service fit-out dates are "expanding from weeks to several months", and conversions from air to liquid cooling face delays that "can extend beyond six months" [28]. NVIDIA's guide puts a DGX GB rack, a 72-GPU system, at approximately 120 kW [22]. Rack power and liquid cooling covers what that load requires.
Two delays, two stages
| CoreWeave | SANGRIX | |
|---|---|---|
| What slipped | Powered shell from a third-party developer | Delivery of 16 B300 servers |
| Timing | Delays affected fourth-quarter 2025 expectations, with most of the delay expected to clear within the first quarter of 2026 | Third quarter of 2026, moved to the week of 3 November 2026, plus about a week for inspection, racking and testing |
CoreWeave's chief executive said in November 2025 that "really where the challenge is, is the powered shell" [17]. Construction in progress rose by $2.8 billion in the third quarter of 2025, which the company tied to the delays [17]. The customer agreed to adjust the delivery schedule and extend the expiration date, and CoreWeave kept the contract's total value [17]. SANGRIX's new date came from its seller, relaying an unnamed manufacturer, with no reason given [18]. It is one small company's disclosure.
Rillor offers forward agreements that fix the grade, quantity, price and delivery month of GPU systems at signing, and capacity agreements that reserve power and space in fixed blocks before the hardware arrives. Forward agreements for GPU systems explains how such an agreement works.
Read the full report
The full report sets out each stage's statements with their scope and dates, the complete supplier capacity timeline, CBRE's market figures and both delays in detail, with all 34 sources.