Compiled by XAIPART Editorial from public sources through 2026-11-17. Facts carry numbered citations; forward-looking judgments are labeled XAIPART analysis.
Scope. This report covers AI accelerators — data-center GPUs, custom ASICs, and the HBM and advanced packaging around them — as of November 17, 2026. It draws on vendor newsrooms, regulatory filings, research firms, and international trade press. Every factual sentence carries a numbered source; forward-looking judgments are labeled XAIPART analysis.
1. Executive summary
On November 16, 2026, NVIDIA used SC26 to announce Rubin HPC: a single GPU with more than 50,000 CUDA cores, seventh-generation NVLink, and a design aimed at simulation and data analytics, with availability in 2028 S1. The same week, the 68th TOP500 list counted 291 accelerator-based systems, or 58.2% of the list; the top ten did not change, and China's LineShine remained first at 2.198 exaflops S2. On November 12, AMD expanded its MI400 family: MI450X carries 288GB of HBM4, MI430X carries 432GB and targets scientific computing, and MI455X reached 21,333 tokens per second per GPU on DeepSeek-R1 in an AMD-tested MLPerf Inference Offline configuration. AMD says that result is up to 30% above NVIDIA's published Vera Rubin NVL72 figure S3.
Demand signals diverge. SemiAnalysis cut its cumulative 2026–2030 forecast for accelerator logic wafer shipment area by 17% and reduced its 2028 HBM shipment forecast from 9.0 EB to 6.4 EB S5. TrendForce expects HBM consumption to grow more than 70% in 2026 S6. BofA values the 2026 HBM market at $54.6 billion, up 58% S7. The three measures count different things — wafer area, HBM bits, and dollars — over different time windows.
XAIPART analysis: for a component distributor, the addressable business is not the GPU or the HBM stack. It is the power, passive, connector, and optical content of an AI rack, where lead times and allocation are already tighter than prices. Before 2027, allocation capability matters more than price forecasting.
Key numbers at a glance
| Metric | Value | Source |
|---|---|---|
| Rubin HPC CUDA cores | More than 50,000 per GPU | S1 |
| TOP500 accelerator systems | 291 of 500 (58.2%) | S2 |
| AMD MI455X DeepSeek-R1 | 21,333 tokens/s/GPU | S3 |
| SemiAnalysis wafer-area cut | -17% cumulative, 2026–2030 | S5 |
| TrendForce 2026 HBM consumption | More than +70% YoY | S6 |
| BofA 2026 HBM market | $54.6 billion, +58% | S7 |
| AI-server high-cap MLCC lead time | 26–40 weeks; some 52 weeks | S43, S44 |
How to read this report. Facts carry numbered citations. Passages labeled XAIPART analysis are our own judgments. When sources disagree, the conflict is recorded rather than resolved.
2. Historical development
Timeline: from CUDA to Rubin HPC
- CUDA ships with G80, turning GPUs into programmable parallel processors. S19
- AlexNet trains on two GTX 580 cards and wins ImageNet, validating GPU deep learning. S20
- Google deploys TPU v1 for inference; AMD ships the first commercial HBM1 in Radeon R9 Fury X. S63S37
- Volta V100 introduces the first Tensor Cores and NVLink 2. S65
- A100 80GB brings HBM2e and doubles per-card memory. S64
- H100 adds Transformer Engine and FP8. S18
- Blackwell GB200 NVL72 moves the unit of competition from card to rack. S16
- Blackwell Ultra GB300 NVL72 ships; SK hynix completes HBM4 development. S17S36
- Vera Rubin enters full production; Rubin HPC is announced at SC26; Samsung ships commercial HBM4 first. S15S1S34
- Rubin Ultra NVL576 targets 576 dies, about 600kW, and 800VDC distribution. S66S39
- Rubin HPC becomes available for scientific computing. S1
2.1 Four eras
| Era | Period | Key products and events | Demand characteristics | Sources |
|---|---|---|---|---|
| GPGPU and academic computing | 2006–2012 | CUDA launched with G80 in 2006; Tesla compute cards; AlexNet trained on two GTX 580s in 2012 | Research and a small number of internet recommendation workloads; GPUs move from graphics to general parallel computing | S19, S20 |
| Deep-learning takeoff | 2012–2017 | Kepler (2012), Maxwell (2014), Pascal P100 (2016), Volta V100 (2017, first Tensor Cores) | Training clusters grow from tens of cards to hundreds; HBM enters the data center | S19, S65 |
| Cloud data-center scale | 2017–2022 | A100 (2020, HBM2e), H100 (2022, Transformer Engine), TPU v4 (2021), Trainium (2020) | Hyperscale purchasing, cloud rental, model parameters from millions to hundreds of billions | S18, S64, S30 |
| Generative AI, rack scale, and ASIC rise | 2022–present | H100/H200, Blackwell NVL72 (2024), Blackwell Ultra (2025), Vera Rubin (2026), Rubin Ultra (2027), Rubin HPC (2028); TPU v7/v8, Trainium3, Maia 200 | Rack power rises from tens of kW to hundreds of kW; training and inference split; custom ASIC share increases | S15, S16, S17, S21, S22, S24, S26, S1 |
CUDA was a niche programming model when NVIDIA shipped it in 2006. In 2012, AlexNet trained on two 3GB GTX 580 cards for five to six days, won ImageNet, and gave GPU deep learning a reproducible cost curve S20. Volta V100 moved matrix multiply-add into Tensor Cores in 2017 S65. A100 80GB added HBM2e in 2020, doubling per-card memory and making model parallelism affordable S64. H100 added Transformer Engine and FP8 in 2022, giving attention layers a hardware path for precision scheduling S18.
Blackwell changed the unit of competition from the card to the rack in 2024: GB200 NVL72 combines 72 Blackwell GPUs, 36 Grace CPUs, fifth-generation NVLink, and liquid cooling in one rack-scale system S16. Blackwell Ultra's GB300 NVL72 delivers 1.5 times the AI performance of GB200 NVL72 S17. NVIDIA said Vera Rubin entered full production on May 31, 2026 S15. At SC26 it added Rubin HPC for scientific computing, due in 2028 S1. The next rack-scale step is Rubin Ultra NVL576 in 2027: 576 GPU dies, about 600kW per rack, 15 exaflops of FP4, 800VDC distribution, and separate power and cooling sidecars S66.
Rack scale binds interconnect, power, cooling, switching, and storage into one delivery unit. Cloud buyers purchase usable rack capacity, not a standalone card. GB200 NVL72 connects 72 GPUs into one NVLink domain; GB300 NVL72 raises performance by another 1.5 times S16S17. AMD's Helios uses MI455X and MI450X GPUs in the same rack architecture S3. XAIPART analysis: once the unit is a rack, the barrier for a second source shifts from "the chip runs" to "the rack ships," which is why ASIC vendors must coordinate with Broadcom, ODMs, and liquid-cooling suppliers.
Custom ASICs started earlier. Google deployed TPU v1 in 2015 and disclosed it in 2016; the first version handled inference with 8-bit integer arithmetic, and Google said it was 15 to 30 times faster and 30 to 80 times more energy-efficient than contemporary CPUs and GPUs S63. The line later split into training and inference: TPU v7 Ironwood (2025) delivers 4,614 TFLOPS of FP8, 192GB of HBM3e, and 7.37TB/s per chip, with 9,216 chips forming a 42.5-exaflop pod S21. On April 22, 2026, Google announced its eighth-generation TPU, splitting training (TPU 8t) and inference (TPU 8i) into two chips, with Broadcom and MediaTek sharing the work S22. AWS moved from Inferentia in 2019 to Trainium3, generally available in December 2025: a 3nm chip, 144 chips per UltraServer, 4.4 times the compute of Trainium2, and 20.7TB of HBM3e S24. Microsoft announced Maia 200 on January 26, 2026: TSMC 3nm, 216GB of HBM3e, 7TB/s, 272MB of on-chip SRAM, more than 10 petaflops of FP4, and a 750W envelope S26. Meta outlined four MTIA generations in March 2026 — MTIA 300, 400, 450, and 500 — with MTIA 300 already in production for ranking and recommendation training S27.
Memory followed the same cadence. AMD's Radeon R9 Fury X shipped the first commercial HBM1 in 2015: 4GB, a 4096-bit interface, and 512GB/s S37. SK hynix mass-produced HBM2E in 2020, developed HBM3 in October 2021, mass-produced 12-layer HBM3E in September 2024, and completed HBM4 development in September 2025 S36. Samsung shipped the industry's first commercial HBM4 on February 12, 2026, and sent 12-layer HBM4E samples on May 29, 2026 S34S35. On the packaging side, TSMC's CoWoS capacity is moving from about 70,000 wafers per month in 2025 to 120,000–140,000 by the end of 2026 S32.
2.2 NVIDIA iteration map
| Architecture | Year | Representative product | Key features | HBM generation | Sources |
|---|---|---|---|---|---|
| Tesla | 2007 | Tesla C870 | First GPU compute cards; CUDA available | GDDR3 | S19 |
| Fermi | 2010 | Tesla M2050 | ECC and a deeper double-precision cache hierarchy | GDDR5 | S19 |
| Kepler | 2012 | Tesla K20X | Dynamic parallelism and GPUDirect | GDDR5 | S19 |
| Maxwell | 2014 (architecture) / 2015 (M40) | Tesla M40 | Higher performance per watt; inference begins | GDDR5 | S19 |
| Pascal | 2016 | Tesla P100 | First data-center GPU with HBM and NVLink 1 | HBM2 | S65 |
| Volta | 2017 | Tesla V100 | First Tensor Cores and NVLink 2 | HBM2 | S65 |
| Turing | 2018 | Tesla T4 | INT8/FP16 inference optimization and low power | GDDR6 | S18 |
| Ampere | 2020 | A100 40/80GB | TF32, MIG, and NVLink 3 | HBM2e | S64 |
| Hopper | 2022 | H100/H200 | Transformer Engine, FP8, and NVLink 4 | HBM3/HBM3e | S18 |
| Blackwell | 2024 | B200/GB200 NVL72 | Dual die, fifth-generation NVLink, liquid-cooled rack | HBM3e | S16 |
| Blackwell Ultra | 2025 | GB300 NVL72 | 1.5 times GB200 NVL72 | HBM3e | S17 |
| Vera Rubin | 2026 | Vera Rubin NVL72 | Full production, HBM4, rack scale | HBM4 | S15 |
| Rubin Ultra | 2027 | NVL576/Kyber | 576 dies, about 600kW, 800VDC | HBM4 | S66 |
| Rubin HPC | 2028 | Rubin HPC | More than 50,000 CUDA cores and seventh-generation NVLink | Not disclosed | S1 |
The architecture years for 2007–2016 follow NVIDIA's corporate timeline; representative card names are public product names, and their detailed specifications were not rechecked line by line, so no memory or bandwidth figures are given. The 600kW and 576-die figures for Rubin Ultra come from a third-party summary of NVIDIA material and remain single-source S66.
2.3 Custom ASIC map
| Vendor | Generation/product | Year | Key specification or event | Sources |
|---|---|---|---|---|
| TPU v1 | Deployed 2015, disclosed 2016 | Inference ASIC, 8-bit integer, 15–30 times faster than CPU/GPU | S63 | |
| TPU v2 | 2017 | Training and inference, HBM, pod interconnect | S78 | |
| TPU v3 | 2018 | Liquid cooling and larger pods | S78 | |
| TPU v4 | 2021 | Optical circuit switching and large-scale training | S78 | |
| TPU v5e/v5p | 2023 | Separate inference and training tiers | S78 | |
| TPU v6e Trillium | 2024 | Sixth generation and higher efficiency | S78 | |
| TPU v7 Ironwood | 2025 | 4,614 TFLOPS FP8, 192GB HBM3e, 9,216 chips/42.5 EF | S21 | |
| TPU 8t/8i | 2026 | Training/inference split; Broadcom and MediaTek; general availability in the second half of 2026 | S22, S23 | |
| AWS | Inferentia/Inf1 | 2019 | Inference ASIC | S73 |
| AWS | Trainium/Trn1 | Announced 2020; software support in 2022 | Training ASIC; general-availability month not verified | S73 |
| AWS | Trainium2/Trn2 | 2024 | Training and inference, UltraServer | S25 |
| AWS | Trainium3/Trn3 | GA December 2025 | 3nm, 144 chips per rack, 4.4 times, 20.7TB HBM3e | S24 |
| Meta | MTIA v1 | 2023 | Inference; specifications not disclosed | S74 |
| Meta | MTIA v2 | 2024 | Inference; specifications not disclosed | S74 |
| Meta | MTIA 300 | 2026 | Ranking and recommendation training; in production | S27 |
| Meta | MTIA 400/450/500 | Announced 2026; deployment in 2027 | A training-and-inference portfolio | S27 |
| Microsoft | Maia 100 | 2023 | Training and inference; specifications not disclosed | S26 |
| Microsoft | Maia 200 | January 2026 | TSMC 3nm, 216GB HBM3e, 7TB/s, more than 10 PFLOPS FP4, 750W | S26 |
| Intel | Gaudi 2 | 2022 | Discontinued | S75 |
| Intel | Gaudi 3 | 2024 | Last discrete AI accelerator in the line | S75, S28 |
| Intel | Falcon Shores | Commercial launch canceled in 2026 | Internal test chip only | S75 |
| Intel | Jaguar Shores | 2027 | Rack-scale system | S75 |
| Intel | Crescent Island | Customer samples in the second half of 2026 | Data-center GPU with LPDDR5X | S28 |
The TPU v2–v6 years follow Google's public generation history and an academic review, without per-generation performance figures S78. Meta MTIA v1/v2 and Microsoft Maia 100 did not disclose complete specifications in public material, so the table says "not disclosed" rather than estimating S26S27S74.
2.4 AMD Instinct map
| Generation | Year | Representative product | Key features | HBM generation | Sources |
|---|---|---|---|---|---|
| MI200 | 2021 | MI250X | CDNA2, chiplet design, CPU+GPU supercomputers | HBM2e | S76 |
| MI300 | 2023 | MI300X/MI300A | CDNA3, chiplet and HBM3 co-packaging | HBM3 | S76 |
| MI325X | October 10, 2024 | MI325X | Higher memory capacity and bandwidth | HBM3e | S29, S76 |
| MI350 | 2025 | MI350X | CDNA4; AMD claims up to 35 times the inference performance of MI300 | HBM3e | S30 |
| MI400 | 2026 | MI430X/MI450X/MI455X | Next CDNA generation, HBM4, Helios rack | HBM4 | S3 |
AMD split the MI400 family into three tiers on November 12, 2026. MI430X targets scientific computing with 432GB of HBM4, 23.3TB/s, and native FP64. MI450X targets AI with 288GB of HBM4 and the same core compute as MI455X. MI455X is the flagship for Helios S3. AMD also published a DeepSeek-R1 MLPerf Inference Offline result of 21,333 tokens per second per GPU for MI455X and said it is up to 30% above NVIDIA's published Vera Rubin NVL72 entry. The comparison uses an AMD-tested configuration against NVIDIA's public MLPerf v6.1 result; NVIDIA had not responded at the time of writing, and the figure is a vendor claim S3. XAIPART analysis: AMD is segmenting customers by memory capacity and precision rather than covering every workload with one chip.
3. Demand
3.1 Hyperscaler capex: three ways to total the same quarter
| Company | Q3 2026 capex | Definition | Sources |
|---|---|---|---|
| Amazon | $75.9 billion | Cash capex | S58 |
| Alphabet | $51.4 billion | Cash capex | S58 |
| Microsoft | $45.2 billion | Capex including finance leases | S58 |
| Meta | $34.9 billion | Capex including finance leases | S58 |
| Four-company total | About $207 billion | Mixed definitions above | S58 |
| Four-company total | More than $210 billion, up 22% QoQ and 84% YoY | Platformonomics definition | S57 |
| Four-company total | $171.2 billion | Another statistical definition | S72 |
Three totals — $171.2 billion, $207 billion, and more than $210 billion — describe the same quarter. The difference is definitional: whether finance leases, accrued but unpaid items, land, and power facilities are included; whether the measure is cash spending or accrual accounting. Microsoft separately guides to about $170 billion for the calendar year and said some third-party capacity deliveries expected in Q1 slipped S58. XAIPART analysis: demand tracking must hold the definition fixed. Adding cash capex to capex including leases can overstate or understate the total by 15% to 20%.
3.2 Three forecasts: wafer area, HBM bits, and dollars
Wafer area: -17%
SemiAnalysis counts cumulative 2026–2030 logic silicon area, not unit cards. S5
HBM bits: +70%
TrendForce counts 2026 memory content. Bit growth can coexist with a wafer-area cut. S6
HBM dollars: +58%
BofA counts 2026 revenue, including price and product mix. S7
| Measure | Institution | Window | Figure | What it counts | Sources |
|---|---|---|---|---|---|
| Accelerator logic wafer shipment area | SemiAnalysis | Cumulative 2026–2030 | Cut by 17% | Silicon area, not unit cards | S5 |
| AI accelerator capex | SemiAnalysis | 2028 | $1.054 trillion to $758 billion, -28% | Capex dollars | S5 |
| AI accelerator capex | SemiAnalysis | 2030 | $1.531 trillion to $923 billion, -40% | Capex dollars | S5 |
| HBM shipments | SemiAnalysis | 2028 | 9.0 EB to 6.4 EB, -30% | HBM bits | S5 |
| HBM consumption | TrendForce | 2026 | More than +70% YoY | HBM bits | S6 |
| HBM market size | BofA | 2026 | $54.6 billion, +58% | Dollars | S7 |
| HBM demand for ASICs | Goldman Sachs | 2026 | +82%, one-third of the market | HBM used by ASICs | S7 |
XAIPART analysis: the three figures do not share a denominator. Wafer area proxies logic die supply. HBM bits measure memory content per system. Dollars add price and product mix. HBM3E 12-layer prices rose more than 50% in June 2026, and HBM4 12-layer 36GB prices could double in 2027 S47. The same bit count therefore produces more revenue. A 17% cumulative cut in wafer area, 70% growth in HBM bits, and 58% growth in HBM dollars can all be true at once.
The time window matters more than the sign. SemiAnalysis concentrated its cuts in 2028–2030: capex down 28% in 2028, 41% in 2029, and 40% in 2030, with NVIDIA accounting for 66% of the wafer-area reduction S5. TrendForce and BofA describe 2026 tightness. Read together, the forecasts point to near-term tightness and out-year easing, not an immediate 2026 collapse. XAIPART analysis: treating SemiAnalysis's five-year cumulative cut as a 2026 demand collapse is a category error. Using TrendForce's 2026 +70% to dismiss the 2028 oversupply risk is selective reading.
Server shipments are the fourth measure and the one closest to component distribution. TrendForce expected AI server shipments to grow more than 28% in January 2026 and raised that to nearly 31% in August, citing a 90% increase in 2026 capex across major North American cloud providers and Oracle S10. The same firm cut general-purpose server shipment growth from 20% to 13% in April because PMIC and BMC lead times stretched S45. XAIPART analysis: the two numbers are consistent. AI servers are absorbing capacity and parts that would otherwise go to general servers. IDC's Q2 2026 data gives the dollar view: worldwide server revenue was $166.3 billion, GPU-accelerated servers contributed $87.4 billion (+28.1%), and other accelerated servers, mainly ASICs, contributed $27.5 billion (+237.6%) S55.
SemiAnalysis's public summary describes a specific chain. It cut 2028 global AI accelerator capex from $1.054 trillion to $758 billion, then cut 2029 by 41% and 2030 by 40%. It reduced cumulative 2026–2030 accelerator logic wafer shipment area by 17%, with 66% of the reduction from NVIDIA. It cut 2028 HBM shipments from 9.0 EB to 6.4 EB and lowered its 2030 AI accelerator power demand forecast S5. The cuts concern capex and power availability after 2028, not orders already placed for 2026. The public summary does not disclose customer-level or model-level detail; the underlying work is a paid report.
One transcription conflict is worth recording. SK hynix's newsroom cites BofA's 2026 HBM market forecast at $54.6 billion, up 58% S7. A Chinese-language summary attributed to the same institution put 2026 HBM sales at $26.94 billion. The two differ by roughly a factor of two and the lower figure is inconsistent with the roughly $35 billion scale implied by the 2025 base. XAIPART analysis: the $26.94 billion figure is likely a unit or definition transcription error. This report uses $54.6 billion and lists the lower figure as unverified.
3.3 Inference and training
Inference is becoming a separate demand pool. Google calls Ironwood its first TPU designed for inference S21. Microsoft positions Maia 200 as an inference accelerator S26. AMD uses a DeepSeek-R1 MLPerf Offline result to demonstrate MI455X inference throughput S3. Meta's MTIA 300 starts with ranking and recommendation training, while 400/450/500 extend toward higher-end inference S27. Training remains NVIDIA's stronghold: Vera Rubin's full production and GB300 NVL72's 1.5-times gain point to rack-scale training clusters S15S17, while Rubin CPX adds a dedicated tier for long-context inference S68.
Shipment mix points in the same direction. TrendForce expects about 2.8 million AI servers in 2026, up 30.5%, with GPU servers at 69.7% and ASIC servers at 27.8%, the highest ASIC share since 2023; ASIC shipments are growing faster than GPU shipments S11S56. IDC's Q2 2026 data is more direct: GPU-accelerated server revenue was $87.4 billion (+28.1%), while other accelerated servers, mainly ASICs, reached $27.5 billion (+237.6%) S55. ASICs start from a small base, grow faster, and remain about one-third of GPU revenue.
XAIPART analysis: inference fragmentation favors ASICs. Inference customers segment by token cost, latency, memory capacity, and model size; one GPU SKU cannot cover every tier. A custom ASIC can remove general-purpose capability and trade it for lower cost per token. Training runs the other way: the NVLink domain, CUDA ecosystem, and rack-scale delivery raise migration costs, so training remains NVIDIA's profit pool. The two coexist in the same capex pool at different growth rates.
4. The component ripple
Ripple chain at a glance
Each node passes cost and lead time to the next. Detailed evidence follows in 4.1–4.6.
4.1 Advanced process and packaging
TSMC put N2 into volume production in the second half of 2025; N2P and A16 are scheduled for the second half of 2026, with A16's Super Power Rail aimed at high-density HPC power delivery S31. The binding constraint is packaging, not wafer starts. CoWoS capacity rises from about 70,000 wafers per month in 2025 to 120,000–140,000 by the end of 2026, and the supply-demand gap is expected to narrow from 20% to 10% S32. In October 2026, supply-chain sources said TSMC was considering a 10% to 20% CoWoS price increase, with some packaging lead times already extended S33.
XAIPART analysis: a CoWoS price increase travels along two paths. The first is accelerator ASP, where packaging takes a larger share of bill-of-materials cost and part of the increase passes to cloud buyers. The second is packaging materials and equipment: ABF substrates, interposers, thermocompression bonding, underfill, test sockets, and inspection. XAIPART does not sell CoWoS, but it can sell connectors, probes, fixtures, cleaning, and thermal-control parts around packaging and test. The final size of the price increase has not been confirmed by TSMC; 10% to 20% is a supply-chain figure S33.
4.2 HBM4: the tightest link
Samsung shipped commercial HBM4 first on February 12, 2026, and sent 12-layer HBM4E samples on May 29, 2026 S34S35. SK hynix began HBM4 mass shipments in the second quarter of 2026, with volume expanding in the second half, and sent HBM4E samples to customers in the first half of 2026 S67. Supply remains concentrated. Counterpoint's Q2 2026 estimate is SK hynix 50%, Samsung 33%, and Micron 18% S48. Analyst estimates for NVIDIA's 2026 HBM4 allocation put SK hynix at 60% to 70%, Samsung at 25% to 30%, and Micron at the remainder S49. The two sets disagree, and neither is an NVIDIA disclosure.
Price signals are harder. HBM3E 12-layer prices rose more than 50% in June 2026, and HBM4 12-layer 36GB prices could double in 2027 S47. SK hynix's 2027 HBM4 capacity is reportedly already committed to customers, according to Cantor Fitzgerald, with NVIDIA paying about $32 per GB and Broadcom about $36 per GB S46. XAIPART analysis: HBM allocation is a multi-quarter or multi-year capacity reservation, not a spot purchase. A distributor cannot sell the HBM stack, but it can sell test, packaging, thermal, and power-delivery parts around it. For procurement, HBM is the hardest line in the AI-server BOM to hedge; long-term agreements matter more than spot prices.
4.3 Networking: 800G is the volume, 1.6T is the ramp
800G remains the main optical-module volume in AI data centers in 2026, while 1.6T moves into volume ramp. The tightest parts are 200G-per-lane lasers and DSPs S41. Goldman Sachs sees the optical networking TAM reaching $154 billion in 2028 S40; in September 2026, the bank raised its 2027 1.6T shipment forecast from 42 million units two months earlier to 71.4 million units, as reported by 36Kr S79. NVIDIA uses seventh-generation NVLink on Rubin HPC, and both scale-up inside the rack and scale-out between racks raise optical content S1.
XAIPART analysis: the optical BOM is more fragmented than the GPU BOM. A 1.6T module includes a DSP, driver, TIA, lasers, FAU, connectors, cages, thermal parts, and power management. Most of these part numbers can move through distribution. The risk is qualification: a new 1.6T DSP and 200G-per-lane lasers must be tuned with switches, NICs, and server platforms. Stocking too early ties up inventory; stocking too late misses the ramp.
4.4 Power and cooling: from 54VDC to 800VDC
NVIDIA plans full production of 800VDC data centers with Kyber racks in 2027, targeting 1MW-class racks. A 1MW rack at 54VDC needs about 200kg of copper busbar; 800VDC cuts current and copper losses S39. Rubin Ultra NVL576 draws about 600kW per rack and uses separate power and cooling sidecars S66. Liquid-cooling penetration among AI chips is expected to reach 53% in 2026 and nearly 60% in 2027 S38.
The macro power numbers conflict. A Gartner figure cited in Chinese media puts 2026 data-center electricity use at 565TWh, up 26%, and 2030 above 1,200TWh. An IEA figure cited in another report puts 2026 near 1,000TWh S70S71. The two use different boundaries for "data center," different treatment of on-site generation and network equipment, and different IT-load versus campus-load definitions. This report does not resolve the conflict; it treats the direction as evidence that power is becoming a harder constraint. At least 75 U.S. data-center projects were halted or delayed in 2026, according to media reports that were not individually verified S70.
XAIPART analysis: power and cooling are the most addressable layers for XAIPART. 800VDC shifts value from low-voltage MOSFETs to high-voltage SiC/GaN, solid-state breakers, DC busbars, BBUs and supercapacitors, connectors, and high-voltage MLCCs. Liquid cooling shifts value to CDUs, pumps, quick disconnects, hoses, cold plates, fans, and leak detection. The full 800VDC rack BOM has not been published part by part; the list above is inferred from the architecture and is labeled XAIPART analysis.
4.5 Server ODMs and liquid cooling
Foxconn held about 40% of AI rack assembly in 2026, said AI rack shipments would double, and reported June 2026 revenue of about NT$1.33 trillion S42. Quanta, Wistron, and Hon Hai together cover 80% to 90% of AI server manufacturing demand S41. Rack-level delivery moves purchasing authority from the cloud provider's chip team to the ODM's system team, and the bill of materials moves from board level to rack level.
XAIPART analysis: the ODM is the distributor's real customer, not the cloud provider. The cloud provider specifies the GPU, HBM, and switch silicon; the ODM decides second and third sources for power, connectors, passives, fans, and cables. To capture AI-server demand, XAIPART should break the ODM rack BOM into quotable part numbers instead of stopping at "AI chips."
4.6 What it means for component distribution
| Link | 2026 signal | Distribution implication | Sources |
|---|---|---|---|
| High-capacitance MLCC | Server lead times above 20 weeks, some 24 weeks; high-cap/high-voltage parts 26–40 weeks with 50%–60% price increases | Lock volume with long-term agreements, qualify second sources, avoid chasing spot prices | S43, S44 |
| Standard MLCC | Standard packages at 14–18 weeks | Normal lead time; can buffer a bundled quote | S43 |
| PMIC/BMC | Shortages cut 2026 general-server shipment growth from 20% to 13% | Power-management ICs are the hidden bottleneck in general servers | S45 |
| Memory | HBM capacity essentially sold out for 2027; HBM3E 12-layer prices rose more than 50% in June 2026 | Memory inflation passes into the full system BOM | S46, S47 |
| Optics | 800G is the volume; 1.6T is ramping; 2027 1.6T forecast at 71.4 million units | DSP, driver, TIA, FAU, and connector demand pulls forward | S40, S41 |
| Power | 800VDC production in 2027; 600kW–1MW per rack | SiC/GaN, busbars, BBUs, high-voltage connectors | S39, S66 |
XAIPART analysis: the value of AI servers to a distributor is not one more expensive chip. It is more part numbers per machine, with longer lead times. High-capacitance MLCC lead times stretch from 14–18 weeks for standard parts to 26–40 weeks for AI-server grades, and some reports put Samsung Electro-Mechanics server MLCC lead times at 52 weeks in October 2026 S43S44. Shortages in these parts do not make headlines, but they stop a production line. XAIPART's opportunity is one-stop kitting of long-lead-time parts, not reselling GPUs.
5. Competition
5.1 Share measures
| Measure | NVIDIA | Custom ASIC | AMD | Sources |
|---|---|---|---|---|
| AI processor revenue share (March 2026) | About 85% | Slightly above 10% | Slightly below 5% | S12 |
| AI server shipment share (2026 forecast) | GPU servers 69.7% | ASIC servers 27.8% | — | S11 |
| TOP500 accelerator systems (November 2026) | 110 Hopper GPUs | — | 34 Instinct GPUs | S2, S59 |
| Latest quarterly data-center/AI revenue | $89.0 billion (FY2027 Q2) | Broadcom AI $16.7 billion (FY2026 Q3) | $8.546 billion (Q3 2026) | S13, S8, S4 |
| Year-over-year growth | +117% | +221% | +97% | S13, S8, S4 |
Counting caveat. TOP500 counts complete systems, not chips. One system can use NVIDIA GPUs, AMD CPUs, and NVIDIA networking, so "technology present" and "accelerator-driven" are different numbers. S2S59
The TOP500 figure of 291 accelerator-based systems does not mean NVIDIA powers 291 systems. The official release gives only the accelerator total and lists 110 Hopper GPU systems and 34 AMD Instinct systems S2S59. The NVIDIA SC26 blog reviewed for this report focuses on Rubin HPC and does not give a total TOP500 system count S1S2S59. A pre-conference figure of "252 of 500 driven by NVIDIA accelerators" could not be verified on a primary TOP500 page and is not used. The "34 AMD" figure counts Instinct GPUs, not all 202 systems where AMD technology appears S59S60.
The TOP500 counting method also matters. The list ranks measured HPL performance for complete systems, not chip shipments. The November 2026 list contained only 27 new systems, the lowest turnover in the list's history, which suggests new compute is going into AI factories rather than capability-class supercomputers S2. XAIPART analysis: using TOP500 to measure AI-chip share understates GPU training-cluster shipments and overstates the commercial relevance of supercomputers.
5.2 NVIDIA's moat and the real ASIC progress
NVIDIA's moat has three layers: the software migration cost of CUDA, the scale-up domain of NVLink, and the rack-level delivery cadence of an annual roadmap. In FY2027 Q2, ended July 26, 2026, data-center revenue was $89.0 billion, up 117% year over year; Q3 guidance is $108.0 billion, and it excludes data-center compute revenue from mainland China and Hong Kong S13S14. At that scale, absolute revenue can grow even as share erodes slowly.
ASIC progress is real but concentrated. Broadcom's AI semiconductor revenue reached $16.7 billion in FY2026 Q3, up 221% year over year. The company raised its FY2026 AI revenue outlook to about $58 billion and guided to about $115 billion in FY2027 and $230 billion in FY2028; Anthropic is expected to deploy 5GW of Broadcom custom TPUs in FY2027 S8S9. KGI estimates TPU v8 shipments at 4.02 million units in 2026 and 9.10 million in 2027, up 126% S23. TrendForce expects ASIC servers to reach 27.8% of AI server shipments in 2026, the highest since 2023 S11. A third-party summary puts 2026 ASIC shipment growth at 44.6% versus 16.1% for GPUs, but that pair could not be verified on a primary TrendForce page; the widely circulated "ASIC +64.2%, GPU +43.8%" pair is also unverified and is not used here S69.
XAIPART analysis: the ASIC challenge is concentrated in three to five cloud providers, and the open market is not the main battlefield. Broadcom, Marvell, and Alchip sell custom design to large customers, not a general-purpose substitute. NVIDIA's share will drift down from 85%, but it retains pricing power in training clusters and rack-scale delivery. AMD's opening is inference and double-precision scientific computing; MI430X's 432GB of HBM4 is designed for that S3.
5.3 China
The verifiable China facts sit at the manufacturing and control layers. Huawei's Ascend 910C is manufactured on SMIC's N+2, a 7nm-class node, according to an industry-media summary S77. Digitimes reported in February 2025 that 910C yield was about 40%, the first time the Ascend line turned profitable, with a target of 60% S54. The U.S. Bureau of Industry and Security issued guidance on May 13, 2025, saying Ascend 910B/910C/910D and similar advanced-computing ICs were likely developed or produced in violation of U.S. export controls and that transactions involving them carried enforcement risk S50. NVIDIA's H20 was made subject to an export license in April 2025, and the company took a $4.5 billion charge in the first quarter of fiscal 2026 S51. In November 2026, Reuters reported that about 600,000 NVIDIA chips were stranded in Hong Kong awaiting Chinese import approval; another Reuters exclusive reported that Northwestern Polytechnical University, on the U.S. Entity List, bought H200 chips through an intermediary S52S53.
XAIPART analysis: the short-term effect on global AI chips is supply diversion, not demand disappearance. NVIDIA's guidance already excludes China data-center compute revenue, and the $4.5 billion H20 charge was a one-time clearing event. Ascend yield and capacity determine the speed of domestic substitution, but 40% yield is still far from the 60% target S54. For distributors, the China opportunity is in mature-node parts, power, passives, and packaging materials, not advanced GPUs.
6. Risks and uncertainties
| Risk | Evidence on both sides | What to watch | Sources |
|---|---|---|---|
| Demand forecast divergence | SemiAnalysis cuts 2026–2030 wafer area by 17% and 2028 HBM by 30%; TrendForce sees 2026 HBM +70% and BofA sees dollars +58% | 2027 HBM long-term contract prices, CoWoS gap | S5, S6, S7 |
| HBM4 yield and qualification | Samsung ships HBM4 first; SK hynix starts in Q2; share estimates of 50/33/18 versus 60–70/25–30 conflict | HBM4E qualification progress, NVIDIA allocation | S34, S35, S67, S48, S49 |
| Geopolitics and export controls | H20 license and $4.5 billion charge; BIS guidance on Ascend; 600,000 chips stranded in Hong Kong | License approvals, Hong Kong inventory, Entity List enforcement | S50, S51, S52, S53 |
| Circular deals | NVIDIA invested about $30 billion in OpenAI; AMD agreed to buy up to $5 billion of Anthropic stock alongside more than $20 billion of expected compute supply | Customer concentration, receivables, related-party disclosure | S61, S62 |
| Power and grid | Gartner's 565TWh and IEA's roughly 1,000TWh definitions conflict; 75 projects delayed | Grid interconnection queues, PPA prices, rack delivery delays | S70, S71 |
| 2028 oversupply | SemiAnalysis cuts 2028 HBM by 30% and capex by 28%; HBM is sold out for 2027 and HBM3E 12-layer prices rose more than 50% in June 2026 | HBM spot prices, DRAM contract prices, supplier capacity additions | S5, S46, S47 |
Circular deals matter because they can reinforce a demand signal. NVIDIA's investment in OpenAI and AMD's stake in Anthropic, alongside compute-supply agreements, let customer purchases and supplier investments validate each other on the same set of financial statements S61S62. Morningstar lists these arrangements as a governance and revenue-quality issue ahead of IPOs S61. XAIPART analysis: as long as end-market inference revenue grows faster than capex, circular deals are a financing structure. Once inference revenue lags, the same structure amplifies a downgrade.
7. Falsifiable triggers
Each item below is XAIPART analysis. NVIDIA's FY2027 Q3 guidance is $108.0 billion; item 4 uses $115 billion as the floor, which requires at least 6.5% sequential growth and treats anything below that as a break in acceleration.
| # | Assertion | Window | Threshold that proves it wrong | Trigger to watch |
|---|---|---|---|---|
| 1 | HBM stays tight through 2027 | by 2027-06-30 | TrendForce cuts 2026 HBM consumption growth from +70% to below +40% | Two consecutive cuts, with HBM4 36GB prices turning negative QoQ |
| 2 | CoWoS gap narrows below 10% in 1H 2027 | by Q2 2027 | Monthly capacity below 150,000 wafers, or the gap remains above 15% | TSMC capex cut or lower OSAT outsourcing |
| 3 | ASICs exceed 30% of AI server shipments in 2027 | by Q4 2027 | TrendForce's 2027 ASIC share is below 30% | TPU v8 shipments below 7 million or a slower Trainium3 ramp |
| 4 | NVIDIA revenue accelerates in FY2027 Q4 | by 2027-03-31 | Quarterly revenue below $115 billion | Q4 guidance below $120 billion, or China revenue restored but total still short |
| 5 | AI-server high-cap MLCC tightness eases in 1H 2027 | by 2027-06-30 | Server-grade high-cap MLCC lead times stay above 26 weeks | Supplier capacity additions; AI rack shipment growth below 20% |
| 6 | HBM4 36GB prices rise at least 50% in 2027 | by 2027-12-31 | Full-year average less than 50% above the 2026 average | Samsung and SK hynix both add capacity, or HBM4E ramps early |
How to use this table. Each row is a falsification test, not a forecast to defend. If the threshold is met inside the window, the corresponding claim in this report is wrong.
8. Method and limitations
This report takes November 17, 2026 as its cutoff. It cites 79 sources across 61 independent domains, prioritizing vendor newsrooms, regulatory filings, and research firms. Media sources are used for cross-checking and background. The main figures were checked against the cited page or filing; where a figure came from a public summary of a paid report, the text says so.
The report has four limitations. First, SemiAnalysis's underlying model is paid and only its public summary was available; customer-level and model-level detail could not be checked. Second, HBM4 share estimates differ by source and none is an NVIDIA disclosure. Third, the Rubin Ultra NVL576 power and die counts come from a third-party summary and remain single-source. Fourth, no executive interview was conducted, so strategy judgments are labeled XAIPART analysis.
Source conflicts recorded in this report:
- TOP500: a pre-conference figure of 252 NVIDIA accelerator systems could not be verified; the official release counts 291 accelerator-based systems, 110 Hopper GPU systems, and 34 AMD Instinct systems S2S59.
- ASIC growth: a circulating pair of +64.2% for ASIC servers and +43.8% for GPU servers could not be verified; this report uses TrendForce's 27.8% ASIC share and labels the third-party 44.6%/16.1% pair as unverified S11S69.
- HBM share: Counterpoint's 50/33/18 estimate and analyst allocation estimates of 60–70/25–30 for NVIDIA's HBM4 sourcing differ and remain unresolved S48S49.
- Capex totals: $171.2 billion, $207 billion, and more than $210 billion describe the same quarter under different definitions S57S58S72.
- Data-center power: Gartner's 565TWh and IEA's roughly 1,000TWh for 2026 use different boundaries and remain unresolved S70S71.
متبادل اور lifecycle
| پارٹ | تعلق | لائف سائیکل | بنیاد |
|---|---|---|---|
| — | فائل میں کوئی متبادل نہیں ڈیٹا درکار | — | درکار ہے match_type کسی بھی پیشرو لیبل کے ظاہر ہونے سے پہلے |
“Predecessor” اور “alternative” لیبل صرف ریکارڈ شدہ مماثلت کی قسم، لائف سائیکل اسٹیٹس یا ماخذ متن سے آتے ہیں۔ جہاں کوئی مماثلت ریکارڈ نہ ہو، وہاں کوئی لیبل نہیں دکھایا جاتا۔ ہماری اپنی لسٹنگز مارکیٹ ریفرنس نہیں ہیں۔
مسابقتی منظرنامہ
- 6 XAIPART پر اس زمرے میں لائیو لسٹنگز
- 3 نمائندگی شدہ مینوفیکچررز
- 55 پچھلے 30 دنوں میں ہمارے نیوز آرکائیو میں مذکور مینوفیکچررز
گنتیاں صرف XAIPART کیٹلاگ کی عکاسی کرتی ہیں — یہ مارکیٹ شیئر نہیں ہیں۔
متاثرہ parts
کیٹلاگ میں کوئی مماثلت نہیں۔ یہ کہانی XAIPART پر فی الحال درج کسی پارٹ سے مطابقت نہیں رکھتی۔ ہم مطابقت کا اندازہ نہیں لگاتے، اور کوئی پلیس ہولڈر فہرست نہیں دکھاتے۔
بُعد کے لحاظ سے لاگت کا دباؤ
| Dimension | FRED series ID | Unit | Lag | As of |
|---|---|---|---|---|
| Input cost | PCOPPUSDM, PALUMUSDM, PNICKUSDM | USD/mt | 52 d | |
| Supply/demand | PCU334413334413, IPG3344S | index(1982=100) / index(2017=100) | 46 d | |
| FX & energy | DEXCHUS, DCOILBRENTEU, DHHNGSP | rate / USD/bbl / USD/MMBtu | 5 d |
No free authoritative interface for: gold · silver · rare earth · polysilicon · lithium carbonate · market share — listed rather than approximated.
یہ کیسے تیار ہوا
عمل کا ثبوت — fetch، ماخذ کی درستگی، ماڈل اور گیٹ
| Source fetched | NVIDIA Blog · blogs.nvidia.com · 6,596 chars · fetch status success |
|---|---|
| Source sanity (S1) | ok — passed before any model call (no s1_ block recorded) |
| Time contract | 11 <time datetime> values in this brief — 0 violations (+08:00 or date-only) |
اوپر کی ہر قطار پائپ لائن سے تیار ہوتی ہے، ہاتھ سے نہیں لکھی جاتی۔
تبصرے
پہلا تبصرہ کرنے والے بنیں۔
🔐 براہ کرم سائن ان کریں تبصرہ شائع کرنے کے لیے۔
اکاؤنٹ نہیں ہے؟ سائن اپ کریں