Why did Nvidia become so dominant in AI?
Nvidia's position rests on a stack built over two decades, not on any one chip. CUDA in 2006 gave the GPU a C compiler for general computation; AlexNet in 2012 showed that an efficient GPU implementation could train a deep network that beat the field; and Volta's Tensor Cores in 2017 and the A100 in 2020 built the data-center GPU around that workload. H100 began shipping two weeks before ChatGPT's November 2022 release turned generative AI into a capital-expenditure race. By August 2023, data-center revenue had reached $10.32 billion, up 171 percent from a year earlier, as H100 systems shipped to cloud providers and consumer-internet companies; Nvidia then reported fiscal-2024 data-center revenue of $47.5 billion, up 217 percent. The lead also rests on the system around the chip: Mellanox networking, Blackwell NVL72 production while cloud GPUs were sold out, and Vera Rubin NVL72 racks shipping through Supermicro and available at CoreWeave for Cognition production workloads by September 2026. Nvidia's revised HBM4 requirement also forced suppliers to modify their designs. The counterforces are meaningful but mix commitments with alternatives already shipping: AMD and OpenAI agreed on a 6-gigawatt deployment beginning with 1 gigawatt planned for the second half of 2026, and AMD said in October 2026 that it had begun shipping MI455X accelerators and Helios rack-scale systems; Anthropic's Trainium capacity is a multi-year AWS build-out; and Marvell said it was shipping custom silicon to all four U.S. hyperscalers.
Trace the chain
The stack
- 1CUDALaunch
Nvidia unveils CUDA: the GPU becomes a general-purpose processor with a C compiler
Nvidia introduced CUDA as a new architecture for computing on its graphics processors and the first C-compiler development environment for the GPU, pitching hundreds of on-chip cores working in parallel at up to 100 times the speed of conventional approaches. It is the software layer that later let researchers train neural networks on gaming cards, and the ecosystem rivals still have to work around.
NVIDIA press release (Wayback Machine) ·
NVIDIA Corporation (Nasdaq: NVDA), the worldwide leader in graphics processors, today unveiled NVIDIA CUDA technology, a fundamentally new architecture for computing on NVIDIA graphics processing units (GPUs), and the…
Why it matters here. CUDA's C compiler made general-purpose computation on the GPU possible, which is what AlexNet's efficient GPU implementation would run on.
How Nvidia Came to Own AI Compute: From CUDA to the H100 Shortage
- 2AlexNetReport
AlexNet: a deep network trained on two GPUs wins ImageNet by a wide margin
Krizhevsky, Sutskever and Hinton reported a 60-million-parameter convolutional network that cut ImageNet's top-5 error to 18.9 percent, far below the previous state of the art, and credited "a very efficient GPU implementation" for making the training feasible. The result is the moment deep learning and Nvidia GPUs became tied together in research practice.
NeurIPS 2012 proceedings ·
On the test data, we achieved top-1 and top-5 error rates of 39.7% and 18.9% which is considerably better than the previous state-of-the-art results. The neural network, which has 60 million parameters and 500,000…
Why it matters here. An efficient GPU implementation made training a 60-million-parameter network feasible and won ImageNet by a wide margin, creating the training workload that Tensor Cores were designed to accelerate.
How Nvidia Came to Own AI Compute: From CUDA to the H100 Shortage
- 3Tensor CoresLaunch
Volta V100 adds Tensor Cores: the first data-center GPU designed around AI training
At GTC 2017 Nvidia launched the Volta platform and the Tesla V100, with 640 Tensor Cores delivering 120 teraflops of deep-learning performance, which it equated to 100 CPUs. Tensor Cores made matrix math a first-class hardware feature, the design choice every later generation has extended.
NVIDIA Newsroom ·
Equipped with 640 Tensor Cores, V100 delivers 120 teraflops of deep learning performance, equivalent to the performance of 100 CPUs.
Why it matters here. Volta put 640 Tensor Cores for deep learning into the data-center GPU, the design the A100 extended for training and inference.
How Nvidia Came to Own AI Compute: From CUDA to the H100 Shortage
- 4A100Launch
A100 in full production: Ampere unifies training and inference, adopted by every major cloud
Nvidia said the first Ampere GPU, the A100, was in full production and shipping worldwide, with up to 20 times the performance of its predecessors and adoption by the top cloud providers and server makers. The A100 was the chip the large language models of 2020 to 2022 were trained on, which is why the H100 launched into an installed base that already ran on Nvidia.
NVIDIA Newsroom ·
NVIDIA today announced that the first GPU based on the NVIDIA® Ampere architecture, the NVIDIA A100, is in full production and shipping to customers worldwide.
Why it matters here. A100's adoption by top cloud providers and server makers meant that, when H100 began shipping two years later, it entered an installed base already running on Nvidia hardware.
How Nvidia Came to Own AI Compute: From CUDA to the H100 Shortage
- 5H100Shipment
NVIDIA begins shipping Hopper-based H100 GPUs
NVIDIA said it began shipping the Hopper-based H100 Tensor Core GPU. It said the first H100 systems were available.
NVIDIA Newsroom ·
Began shipping the NVIDIA® H100 Tensor Core GPU based on the new NVIDIA Hopper™ architecture, with first systems available now.
Why it matters here. H100 began shipping two weeks before ChatGPT launched, placing Nvidia's new Hopper GPU in the market as generative-AI demand accelerated.
The demand shock
- 6Hyperscaler demandLaunch
OpenAI launches ChatGPT as a free research preview
OpenAI released ChatGPT, a conversational model trained with reinforcement learning from human feedback, free to use during its research preview. Its adoption over the following weeks turned generative AI from a research budget into a capital-expenditure race among cloud providers, all of whom were buying the same GPU.
OpenAI ·
We’ve trained a model called ChatGPT which interacts in a conversational way.
Why it matters here. ChatGPT's free research preview turned generative AI into a capital-expenditure race, and Nvidia's next results reported surging demand.
How Nvidia Came to Own AI Compute: From CUDA to the H100 Shortage
- 7The shortageFinancial
Nvidia reports "surging demand" and says it is significantly increasing supply
With first-quarter fiscal 2024 data-center revenue at a record $4.28 billion, Jensen Huang said the computer industry was going through two transitions at once, accelerated computing and generative AI, and that Nvidia was significantly increasing supply to meet surging demand. The guidance that followed became the clearest public marker of the H100 shortage.
NVIDIA Newsroom ·
“The computer industry is going through two simultaneous transitions — accelerated computing and generative AI,” said Jensen Huang, founder and CEO of NVIDIA.
Why it matters here. Nvidia said it was significantly increasing supply to meet surging demand, the demand signal behind the much larger data-center revenue reported in the next step.
How Nvidia Came to Own AI Compute: From CUDA to the H100 Shortage
- 8H100 revenue surgeFinancial
Data-center revenue of $10.32 billion, up 171 percent in a year
Second-quarter fiscal 2024 data-center revenue more than doubled from the prior quarter to a record $10.32 billion, up 141 percent sequentially and 171 percent from a year earlier, as H100 systems shipped to cloud providers and consumer-internet companies. It was the quarter the demand shift first showed in Nvidia's numbers.
NVIDIA Newsroom ·
Second-quarter revenue was a record $10.32 billion, up 141% from the previous quarter and up 171% from a year ago.
Why it matters here. H100 systems were shipping to cloud providers and consumer-internet companies as data-center revenue reached $10.32 billion, so the full-year results next record the surge at a larger scale.
How Nvidia Came to Own AI Compute: From CUDA to the H100 Shortage
- 9Revenue triplesFinancial
Fiscal 2024 closes with data-center revenue of $47.5 billion, up 217 percent
Nvidia reported fourth-quarter data-center revenue of $18.4 billion, up 409 percent from a year earlier, and full-year data-center revenue of $47.5 billion, up 217 percent. In one fiscal year the data-center business went from a large segment to most of the company, on the back of a single product generation.
NVIDIA Newsroom ·
Fourth-quarter revenue was a record $18.4 billion, up 27% from the previous quarter and up 409% from a year ago. Full-year revenue rose 217% to a record $47.5 billion.
Why it matters here. Data-center revenue of $47.5 billion, up 217 percent in one fiscal year, is the scale at which Blackwell, announced the next month, was introduced.
How Nvidia Came to Own AI Compute: From CUDA to the H100 Shortage
- 10BlackwellAnnouncement
Blackwell platform announced at GTC 2024 with B200 and GB200
Nvidia announced the Blackwell architecture in San Jose, listing the B200 Tensor Core GPU and the GB200 Grace Blackwell Superchip, which pairs two B200 GPUs with a Grace CPU. The release claims up to 25x lower cost and energy for LLM inference compared with H100.
Nvidia Investor Relations ·
Packed with 208 billion transistors, Blackwell-architecture GPUs are manufactured
Why it matters here. Mellanox had already added InfiniBand, Spectrum Ethernet and BlueField to Nvidia’s data-center business before Nvidia announced the Blackwell platform and its GB200 system.
What keeps the leadRuns from 2020, alongside the strands above, not after them
- 11NetworkingAnnouncement
Mellanox deal closes at $7 billion, the start of Nvidia's networking business
The acquisition completed at a transaction value of about $7 billion, adding Mellanox's roughly 3,000 employees and the InfiniBand, Spectrum Ethernet and BlueField DPU lines that became the networking half of Nvidia's data center segment. Within five years networking was a multi-billion-dollar quarterly business inside Nvidia.
NVIDIA Newsroom ·
NVIDIA today announced the completion of its acquisition of Mellanox Technologies, Ltd., for a transaction value of $7 billion. The acquisition, initially announced on March 11, 2019, unites two of the world’s leading…
Why it matters here. Mellanox added InfiniBand, Spectrum Ethernet and BlueField to Nvidia's data-center business; Nvidia later reported Blackwell NVL72 in full-scale production across system makers and cloud service providers.
Nvidia's Dealmaking: From Mellanox and the Failed Arm Bid to Groq and Hugging Face
- 12Rack scaleShipment
NVIDIA says Blackwell NVL72 is in full-scale production
NVIDIA said its Blackwell NVL72 AI supercomputer was in full-scale production. It said production covered system makers and cloud service providers.
NVIDIA ·
Our breakthrough Blackwell NVL72 AI supercomputer — a ‘thinking machine’ designed for reasoning— is now in full-scale production across system makers and cloud service providers.
Why it matters here. Blackwell NVL72 went into full-scale production across system makers and cloud providers, and the next step records those GPUs selling out.
- 13Sold outFinancial
NVIDIA says Blackwell sales and cloud GPUs are sold out
NVIDIA reported Blackwell sales. It said cloud GPUs were sold out.
NVIDIA ·
Blackwell sales are off the charts, and cloud GPUs are sold out.
Why it matters here. Blackwell sales and cloud GPUs were sold out, while the following HBM4 report says strong Blackwell demand adjusted Nvidia's Rubin timeline and its revised specification forced suppliers to modify their designs.
- 14HBMReport
TrendForce: HBM4 mass production slips to the end of the first quarter of 2026 after Nvidia raised the spec
TrendForce said HBM4 mass production was now expected no earlier than the end of the first quarter of 2026, because Nvidia had revised the HBM4 specification for Rubin in the third quarter of 2025 to a per-pin speed above 11 Gbps, forcing all three suppliers to modify designs, and because strong Blackwell demand led Nvidia to adjust the Rubin production timeline.
TrendForce ·
Nvidia has revised the HBM4 specifications for its Rubin platform in 3Q25, raising the required per-pin speed to above 11 Gbps. This change has necessitated that the three major HBM suppliers modify their designs.
Why it matters here. TrendForce reported that Nvidia's revised HBM4 requirement forced suppliers to modify their designs and that strong Blackwell demand adjusted the Rubin timeline; by September a partner was shipping the Rubin rack platform.
HBM Supply Race: SK hynix, Samsung and Micron from HBM3E to HBM4E
- 15Rubin racks shipShipment
Supermicro says it is shipping Vera Rubin NVL72 racks
Supermicro said it was shipping NVIDIA Vera Rubin NVL72 racks. It said the racks were integrated with its Data Center Building Block Solutions and DLC-2 direct-liquid-cooling stack.
Supermicro ·
“Super Micro Computer, Inc. (NASDAQ: SMCI),an AI, Enterprise, Storage, and 5G/Edge IT Total Solution Provider, is now shipping NVIDIA Vera Rubin NVL72 racks integrated with Supermicro's Data Center Building Block…
Why it matters here. Supermicro was shipping NVIDIA Vera Rubin NVL72 racks, and the next step records the system becoming available on a cloud for production workloads.
- 16Cloud production useLaunch
CoreWeave makes Vera Rubin NVL72 available with Cognition workloads
CoreWeave announced the availability of NVIDIA Vera Rubin NVL72 on its cloud, with Cognition as the first customer anywhere running production workloads on the system.
CoreWeave ·
CoreWeave Inc. (Nasdaq: CRWV), The Essential Cloud for AI™, today announced the availability of NVIDIA Vera Rubin NVL72 on CoreWeave, with Cognition as the first customer anywhere running production workloads on the…
Why it matters here. CoreWeave made Vera Rubin NVL72 available with Cognition running production workloads, before the next strand records customers’ commitments to alternative silicon.
What pushes back
- 17AMD competition
OpenAI agrees to deploy 6 gigawatts of AMD GPUs; AMD issues a warrant for up to 160 million shares
AMD and OpenAI announced a multi-year agreement under which OpenAI will deploy 6 gigawatts of AMD Instinct GPUs across generations, starting with 1 gigawatt of MI450 GPUs in the second half of 2026. AMD issued OpenAI a warrant for up to 160 million shares at US$0.01 that vests with deployment milestones and AMD share-price targets.
AMD Investor Relations ·
The first 1 gigawatt deployment of AMD Instinct MI450 GPUs is set to begin in the second half of 2026.
What it pushes against. OpenAI committed to deploy 6 gigawatts of AMD Instinct GPUs, but the first 1-gigawatt MI450 deployment was only set to begin in the second half of 2026.
- 18Hyperscalers' own chips
Amazon puts up to $25 billion more into Anthropic; Anthropic commits $100 billion to AWS and up to 5 GW of Trainium
Amazon invested $5 billion immediately with up to $20 billion more tied to milestones, on top of $8 billion since 2023, and Anthropic committed more than $100 billion over ten years to AWS, securing up to 5 GW across Trainium2, 3 and 4 with nearly 1 GW of Trainium2 and Trainium3 online by end-2026. Anthropic said it already used more than one million Trainium2 chips and kept AWS as its primary training provider.
Anthropic ·
We have signed a new agreement with Amazon that will deepen our existing partnership and secure up to 5 gigawatts (GW) of capacity for training and deploying Claude, including new Trainium2 capacity coming online in…
What it pushes against. Anthropic committed more than $100 billion to AWS and up to 5 GW of Trainium capacity, a multi-year build-out rather than a switch away from Nvidia.
Custom AI Silicon at the Hyperscalers: TPU, Trainium, Maia, MTIA and the OpenAI–Broadcom Chip
- 19Helios begins shipping
Lisa Su says MI455X and Helios shipments have already begun
Speaking in Korea on cooperation with Samsung and SK hynix, Lisa Su said AMD had already begun shipping its next-generation MI455X accelerators and Helios rack-scale systems. It is the first statement that the generation behind the OpenAI, Meta and Anthropic gigawatt agreements has started to leave the factory, rather than being tested or scheduled.
Inven Global ·
We have already begun shipments of our next-generation MI455X and AMD Helios solutions.
What it pushes against. AMD's reported MI455X and Helios shipments push against treating AMD's alternative solely as a future commitment.
- 20Hyperscaler custom silicon
Marvell says it ships custom silicon to four U.S. hyperscalers
Marvell Technology, Inc. said it was shipping custom silicon to all four U.S. hyperscalers. It said fiscal-2026 custom silicon was an approximately $1.5 billion business and described XPU attach across networking, memory, storage, infrastructure management, and inference acceleration.
StockAnalysis, transcript of Marvell Investor Day ·
“We have already announced multiple large strategic agreements around these opportunities, and today we are shipping custom silicon to all four U.S. hyperscalers.”
What it pushes against. Marvell's statement that it was shipping custom silicon, including inference acceleration, to all four U.S. hyperscalers pushes against the claim that Nvidia has no alternatives.
Custom AI Silicon at the Hyperscalers: TPU, Trainium, Maia, MTIA and the OpenAI–Broadcom Chip
Changes to this chain
- Helios shipping moved into the pushback strand: the proposal gave it a strand of its own between two pushback steps, so the page printed that heading twice.
- Answer merged by hand: two proposals landed on the same version and each rewrote the answer without the other's addition. This keeps both the August 2023 H100 quarter and the October 2026 Helios shipments.
- Proposal #203 accepted (publish:amd-instinct-ai-gpus): insert after 18: [counter] Helios begins shipping -> amd-instinct-ai-gpus#1925: AMD's reported MI455X and Helios shipments push against treating AMD's alternative solely as a future commitment.
- Proposal #206 accepted (publish:nvidia-datacenter-gpu-roadmap): insert after 7: [cause] H100 revenue surge -> nvidia-ai-compute-origins#1228: H100 systems were shipping to cloud providers and consumer-internet companies as data-center revenue reached $10.32 billion, so the full-year results next record the surge at a larger scale.; answer: Nvidia's position rests on a stack built over two decades, not on any one chip. CUDA in 2006 gave the GPU a C compiler for general computation; AlexNet in 2012 showed that an efficient GPU implementation could train a deep network that beat the field; and Volta's Tensor Cores in 2017 and the A100 in 2020 built the data-center GPU around that workload. H100 began shipping two weeks before ChatGPT's November 2022 release turned generative AI into a capital-expenditure race. By August 2023, data-center revenue had reached $10.32 billion, up 171 percent from a year earlier, as H100 systems shipped to cloud providers and consumer-internet companies; Nvidia then reported fiscal-2024 data-center revenue of $47.5 billion, up 217 percent. The lead also rests on the system around the chip: Mellanox networking, Blackwell NVL72 production while cloud GPUs were sold out, and Vera Rubin NVL72 racks shipping through Supermicro and available at CoreWeave for Cognition production workloads by September 2026. Nvidia's revised HBM4 requirement also forced suppliers to modify their designs. The counterforces are meaningful but include both future commitments and deployed alternatives: AMD and OpenAI agreed on a 6-gigawatt deployment beginning with 1 gigawatt planned for the second half of 2026, Anthropic's Trainium capacity is a multi-year AWS build-out, and Marvell said it was shipping custom silicon to all four U.S. hyperscalers.
- Proposal #197 accepted (publish:hyperscaler-custom-accelerators): insert after 17: [counter] Hyperscaler custom silicon -> hyperscaler-custom-accelerators#1879: Marvell's statement that it was shipping custom silicon, including inference acceleration, to all four U.S. hyperscalers pushes against the claim that Nvidia has no alternatives.; answer: Nvidia's position rests on a stack built over two decades, not on any one chip. CUDA in 2006 gave the GPU a C compiler for general computation; AlexNet in 2012 showed that an efficient GPU implementation could train a deep network that beat the field; and Volta's Tensor Cores in 2017 and the A100 in 2020 built the data-center GPU around that workload. H100 began shipping two weeks before ChatGPT's November 2022 release turned generative AI into a capital-expenditure race, and Nvidia reported fiscal-2024 data-center revenue of $47.5 billion, up 217 percent. The lead also rests on the system around the chip: Mellanox networking, Blackwell NVL72 production while cloud GPUs were sold out, and Vera Rubin NVL72 racks shipping through Supermicro and available at CoreWeave for Cognition production workloads by September 2026. Nvidia's revised HBM4 requirement also forced suppliers to modify their designs. The counterforces are meaningful but include both future commitments and deployed alternatives: AMD and OpenAI agreed on a 6-gigawatt deployment beginning with 1 gigawatt planned for the second half of 2026, Anthropic's Trainium capacity is a multi-year AWS build-out, and Marvell said it was shipping custom silicon to all four U.S. hyperscalers.
- Proposal #184 accepted (publish:nvidia-datacenter-gpu-roadmap): replace step 14: [cause] Rubin racks ship -> nvidia-datacenter-gpu-roadmap#1480: Supermicro was shipping NVIDIA Vera Rubin NVL72 racks, and the next step records the system becoming available on a cloud for production workloads.; insert after 14: [cause] Cloud production use -> nvidia-datacenter-gpu-roadmap#1752: CoreWeave made Vera Rubin NVL72 available with Cognition running production workloads, before the next strand records customers’ commitments to alternative silicon.; answer: Nvidia's position rests on a stack built up over sixteen years, not on any one chip. CUDA (2006) gave the GPU a C compiler for general computation; AlexNet (2012) showed that an efficient GPU implementation could train a deep network that beat the field; Volta's Tensor Cores (2017) and the A100 (2020) then built the data-center GPU around that workload. The H100 shipped two weeks before ChatGPT (November 2022) turned generative AI into a capital-expenditure race, Nvidia said it was increasing supply to meet surging demand, and data-center revenue reached $47.5 billion in fiscal 2024, up 217 percent. What keeps the lead is the system around the chip: the Mellanox networking bought in 2020, the NVL72 rack in full-scale production by 2025 while cloud GPUs stayed sold out, and Vera Rubin NVL72 racks shipping through Supermicro and available at CoreWeave for Cognition production workloads by September 2026. HBM4 memory’s industry ramp also moved to Nvidia's revised requirement. The counterforces are real but later: AMD's 6-gigawatt OpenAI deal starts deploying in the second half of 2026, and Anthropic's Trainium capacity at AWS comes online over several years rather than at once.
- Step 13 why edited by the editor: Handoff repointed after the roadmap-statement step was replaced by the Supermicro shipment.
- Proposal #166 accepted (publish:nvidia-datacenter-gpu-roadmap): replace step 14: [cause] Rubin racks ship -> nvidia-datacenter-gpu-roadmap#1480: Supermicro was shipping NVIDIA Vera Rubin NVL72 racks, providing concrete evidence of Nvidia’s next rack-scale platform before the next strand records customers’ commitments to alternative silicon.; answer: Nvidia's position rests on a stack built up over sixteen years, not on any one chip. CUDA (2006) gave the GPU a C compiler for general computation; AlexNet (2012) showed that an efficient GPU implementation could train a deep network that beat the field; Volta's Tensor Cores (2017) and the A100 (2020) then built the data-center GPU around that workload. The H100 shipped two weeks before ChatGPT (November 2022) turned generative AI into a capital-expenditure race, Nvidia said it was increasing supply to meet surging demand, and data-center revenue reached $47.5 billion in fiscal 2024, up 217 percent. What keeps the lead is the system around the chip: the Mellanox networking bought in 2020, the NVL72 rack in full-scale production by 2025 while cloud GPUs stayed sold out, and Vera Rubin NVL72 racks shipping through Supermicro by September 2026. HBM4 memory’s industry ramp also moved to Nvidia's revised requirement. The counterforces are real but later: AMD's 6-gigawatt OpenAI deal starts deploying in the second half of 2026, and Anthropic's Trainium capacity at AWS comes online over several years rather than at once.
- Step 14 why edited by the editor: Dropped the implication that the statement answered the TrendForce report; neither entry establishes that.
- Proposal #146 accepted (verify:v9): why of step 9 -> Mellanox had already added InfiniBand, Spectrum Ethernet and BlueField to Nvidia’s data-center business before Nvidia announced the Blackwell platform and its GB200 system.; why of step 13 -> TrendForce reported that Nvidia’s revised HBM4 requirement forced suppliers to modify their designs and that strong Blackwell demand adjusted the Rubin timeline; six months later Nvidia said its roadmap was intact.
- Step 14 why edited by the editor: Why-line written for the strand it now sits in.
- Step 13 why edited by the editor: Rewrote the handoff after the roadmap counter was inserted after this step.
- Proposal #125 accepted (publish:nvidia-acquisitions): insert after 13: [counter] Roadmap intact -> nvidia-datacenter-gpu-roadmap#663: NVIDIA's July statement that its roadmap was intact pushes against the preceding report that strong Blackwell demand had adjusted the Rubin timeline.
- Proposal #11 accepted (verify:v5): why of step 10 -> Mellanox added InfiniBand, Spectrum Ethernet and BlueField to Nvidia's data-center business; Nvidia later reported Blackwell NVL72 in full-scale production across system makers and cloud service providers.; why of step 13 -> Nvidia's revised HBM4 requirement forced all three major suppliers to modify their designs and coincided with an adjusted Rubin timeline; the next strand records an AMD deployment commitment whose first gigawatt was only set to begin in the second half of 2026.
- Proposal #3 accepted (verify:v4): why of step 4 -> A100's adoption by top cloud providers and server makers meant that, when H100 began shipping two years later, it entered an installed base already running on Nvidia hardware.; why of step 9 -> Blackwell was announced with the GB200 system, and Mellanox had already added InfiniBand, Spectrum Ethernet and BlueField networking to Nvidia's data-center business.; why of step 12 -> Blackwell sales and cloud GPUs were sold out, while the following HBM4 report says strong Blackwell demand adjusted Nvidia's Rubin timeline and its revised specification forced suppliers to modify their designs.
- Proposal #2 rejected (verify:v4): Filed by an uncalibrated verify pass; superseded by the recalibrated one
- Proposal #1 rejected (verify:v2): Applied through the seed instead: why-lines rewritten, Blackwell bridge added after Revenue triples, HBM step now cites the TrendForce entry on Nvidia's raised requirement
Timelines in this chain
- How Nvidia Came to Own AI Compute: From CUDA to the H100 Shortage
- Nvidia Data Center GPUs: Hopper, Blackwell and Vera Rubin
- Nvidia's Dealmaking: From Mellanox and the Failed Arm Bid to Groq and Hugging Face
- HBM Supply Race: SK hynix, Samsung and Micron from HBM3E to HBM4E
- AMD Instinct: From MI300X to Helios and the Gigawatt Deals
- Custom AI Silicon at the Hyperscalers: TPU, Trainium, Maia, MTIA and the OpenAI–Broadcom Chip
A chain is written by the editor from entries already published on the timelines above, each of which was read against its source before publication. The chain changes when those entries do. Data and terms.