NVIDIA

Jun 2024 · HBM3E qualification

Micron's HBM3E memory was qualified for NVIDIA's Blackwell platform, narrowing the gap with SK hynix as NVIDIA diversified its HBM supplier base to reduce single-supplier risk.

Jan 2026 · HBM4 certification, full contract

Micron was certified as an HBM4 supplier for NVIDIA's Vera Rubin platform alongside SK hynix and Samsung, with its entire 2026 HBM supply fully contracted — reflecting a shift from a secondary to a core, capacity-constrained supplier relationship.

Now

Micron has been certified as an HBM4 memory supplier for NVIDIA's next-generation Vera Rubin AI platform, alongside SK hynix and Samsung, with its entire 2026 HBM supply fully contracted.

Apr 2025 · Ironwood (TPU v7) launches, explicitly targeting inference

Google announced Ironwood (TPU v7), its first TPU explicitly designed for large-scale AI inference, delivering roughly 10x the performance of its predecessor — positioning Google's own silicon as a genuine alternative to NVIDIA GPUs for inference workloads specifically, while NVIDIA GPUs remained dominant for training.

Nov 2025 · Anthropic's record TPU deal signals real diversification away from NVIDIA

Anthropic closed the largest TPU deal in Google's history, committing to hundreds of thousands of Trillium TPUs in 2026 scaling toward one million by 2027 — a concrete sign that at least one major AI lab is shifting meaningful compute away from NVIDIA GPUs toward Google's own silicon, even as NVIDIA continues supplying Google Cloud's external GPU-instance business.

Now

NVIDIA supplies GPUs to Google Cloud, including new Blackwell-powered G4 instances and planned Vera Rubin NVL72 rack-scale systems, even as Google simultaneously develops competing in-house TPU silicon.

Jan 2025

NVIDIA deepened CUDA's software moat through agentic-AI partnerships with Synopsys ($2B investment, AgentEngineer) and Cadence targeting chip-design and physical-AI workflows, reinforcing CUDA's ecosystem lock-in even as rivals pushed custom silicon.

Now

NVIDIA develops CUDA, the software platform that strengthens the ecosystem surrounding its accelerated computing products.

Dec 2024

Reports emerged that Blackwell's initial GB200/B200 server ramp faced overheating, power-consumption, and interconnect issues, pushing TrendForce to project mass production and peak shipments slipping to Q2-Q3 2025 rather than earlier plans.

Apr 2025

Despite the earlier delay reports, GB200 NVL72 systems began reaching customers — WhiteFiber received its first shipment of Quanta-built GB200 Grace Blackwell Superchip systems, signaling volume production had started.

Now

NVIDIA develops the Blackwell GPU architecture for advanced artificial intelligence and accelerated computing workloads.

Jun 2024 · Blackwell ramp

As NVIDIA began ramping its Blackwell GPU architecture, it moved to secure a larger, multi-year share of TSMC's expanding CoWoS packaging capacity to avoid the bottlenecks seen during the Hopper generation.

Jun 2025 · Multi-year capacity lock-in

NVIDIA reserved the majority of TSMC's advanced CoWoS packaging capacity through 2027 to support Blackwell and next-generation Rubin GPUs, reflecting the relationship's shift from opportunistic ordering to long-horizon capacity planning.

Now

NVIDIA depends heavily on TSMC for manufacturing many of its most advanced semiconductor products.

Oct 2025

HBM3E 12-Hi memory began shipping into NVIDIA's B300 systems, with SK hynix holding roughly 62% HBM revenue share, Micron around 21%, and Samsung around 17% as the three qualified suppliers.

Mar 2026

SK hynix shipped the first HBM4 12-Hi samples to NVIDIA for the upcoming Rubin platform, reported as a 6-9 month lead over Samsung and Micron; NVIDIA separately requested next-generation 16-Hi HBM4 samples from all three suppliers by Q4 2026.

Now

Advanced NVIDIA AI accelerators depend on high-bandwidth memory to support data-intensive artificial intelligence workloads.

Jan 2024

NVIDIA moved to secure a larger, multi-year share of TSMC's expanding CoWoS advanced packaging capacity for Blackwell, aiming to avoid the industry-wide packaging bottlenecks that constrained the prior Hopper generation's ramp.

Jan 2025

NVIDIA reserved the majority of TSMC's advanced CoWoS packaging capacity through 2027 to cover both Blackwell and next-generation Rubin GPU production, underscoring packaging (not wafer fabrication alone) as a primary supply constraint.

Now

NVIDIA's advanced AI processors depend on sophisticated packaging technologies that integrate compute and memory components.

Oct 2025 · OpenAI-AMD 6GW deal signals a marquee customer win

AMD's competitive position strengthened materially when it signed a 6-gigawatt, multi-generation GPU supply deal with OpenAI — including a warrant for up to 10% of AMD's shares — its most significant customer win yet against NVIDIA for a frontier AI lab's compute.

Jan 2026 · NVIDIA share projected to ease toward 75%

Industry estimates projected NVIDIA's AI accelerator revenue share easing from its 2024 peak toward roughly 75% by 2026, as AMD's MI350 series (shipped to hyperscalers in Q3 2025) and hyperscalers' own custom ASICs (targeting 10-15% share) both gained ground.

Now

AMD competes with NVIDIA in accelerated computing and artificial intelligence processors.

Jan 2024

NVIDIA's Blackwell (GB200/GB300) architecture began shipping to hyperscalers, though production faced widely reported delays tied to packaging and memory supply constraints.

2025–2026

NVIDIA advanced its next-generation Vera Rubin platform, certifying Micron, SK hynix, and Samsung as HBM4 suppliers with 2026 HBM supply fully contracted ahead of launch.

Now

NVIDIA develops AI accelerators used to train and operate advanced artificial intelligence systems.

Jan 2025

Combined hyperscaler capex nearly doubled to roughly $410 billion as the AI infrastructure buildout accelerated further.

Jan 2026

Combined hyperscaler capex guidance reached roughly $725 billion — Amazon at $200B, Google at $175-185B, Meta at $115-135B, and Microsoft at $110-120B — a roughly 62% increase over 2025, with analysts projecting the big four could exceed $1 trillion combined by 2027.

Now

Demand for NVIDIA computing platforms contributes to investment in AI-focused data-center infrastructure.

Apr 2023 · Early strategic customer and investor

NVIDIA became an early strategic investor and priority GPU supplier to CoreWeave, then a smaller cloud infrastructure provider, helping it scale into one of the largest NVIDIA-GPU-focused "neocloud" providers ahead of the broader 2024-2025 AI infrastructure boom.

Jan 2026 · $2B additional investment, Rubin/Vera integration

NVIDIA invested a further $2 billion in CoreWeave at $87.20 per share to help it add more than 5 gigawatts of AI compute capacity by 2030, expanding their collaboration to include NVIDIA's upcoming Rubin GPU architecture, Bluefield storage, and Vera CPU line — on top of NVIDIA's prior commitment to buy more than $6 billion in services from CoreWeave through 2032.

Now

NVIDIA supplies GPUs to CoreWeave and has invested directly in the company to accelerate its AI infrastructure buildout.

Jan 2024

Microsoft launched its first in-house AI chip, Maia 100, and announced three successors (codenamed Braga, Braga-R, and Clea) planned for 2025-2027, an explicit hedge against depending on NVIDIA for 94%+ of the GPU market.

Oct 2025

Reports indicated Microsoft's follow-on Maia chip (Braga/Maia 200) mass production slipped into 2026 due to design changes and staff turnover, even as Microsoft's CTO publicly pushed to reduce reliance on NVIDIA and AMD GPUs.

Now

NVIDIA supplies large volumes of GPUs to Microsoft Azure to meet hyperscale AI compute demand.

Jan 2024 · Trainium2 deployed at scale for Anthropic's Project Rainier

AWS deployed Trainium2 at scale — including a 400,000-chip cluster ("Project Rainier") for Anthropic — while continuing to run large amounts of NVIDIA GPU capacity for other customers, showing Trainium supplementing rather than replacing NVIDIA within AWS's infrastructure.

Dec 2025 · Trainium3 launches, but roadmap stays NVIDIA-interoperable

AWS launched Trainium3, its first 3nm AI chip, while teasing an NVIDIA-friendly roadmap: the next-generation Trainium4 is planned to support NVIDIA's NVLink Fusion interconnect, allowing Trainium and NVIDIA GPUs to interoperate in the same clusters rather than AWS pursuing full silicon independence from NVIDIA.

Now

NVIDIA supplies GPUs to AWS to meet hyperscale AI compute demand, alongside AWS's own custom silicon efforts.

Jan 2024

Fifth-generation NVLink shipped with the Blackwell platform, enabling up to 72 GPUs to operate as a single coherent fabric within the GB200 NVL72 rack-scale system, a major scale-up from prior generations' smaller GPU pod sizes.

Jan 2026

NVIDIA outlined a roadmap extending NVLink-based rack-scale coherence toward NVL1152 configurations, alongside co-packaged optics, to further scale the number of GPUs operating as a single fabric within future AI factory designs.

Now

NVIDIA develops NVLink as its proprietary GPU-to-GPU interconnect, exclusive to NVIDIA hardware.

Mar 2024

NVIDIA announced the X800 series — Quantum-X800 InfiniBand and Spectrum-X800 Ethernet — the first networking platforms capable of end-to-end 800Gb/s throughput, adopting a two-pronged strategy of InfiniBand for the highest-performance dedicated AI infrastructure and Ethernet for broader AI-optimized data centers.

Jan 2025

InfiniBand switch sales surged as NVIDIA's Blackwell Ultra platform drove 800Gbps networking demand, with InfiniBand retaining the top-performance tier even as Ethernet (Spectrum-X) held overall market leadership for AI back-end networks.

Now

NVIDIA develops its Quantum InfiniBand networking platform following its acquisition of Mellanox.

Dec 2025

NVIDIA and Synopsys announced an expanded strategic partnership integrating Synopsys' AgentEngineer technology with NVIDIA's agentic AI stack (NIM microservices, NeMo Agent Toolkit, Nemotron models) for autonomous EDA and simulation workflows, alongside a $2 billion NVIDIA investment in Synopsys common stock.

Now

NVIDIA depends on Synopsys electronic design automation software to design its advanced chips.

Jan 2026

Cadence expanded its partnership with NVIDIA to combine Cadence's agentic AI-driven EDA and system design tools with NVIDIA's CUDA-X, AI physics, and Omniverse libraries, extending the relationship into physical AI systems and hyperscale AI factory design.

Now

NVIDIA depends on Cadence electronic design automation software to design its advanced chips.

Jan 2025

NVIDIA announced Spectrum-X Photonics and Quantum-X Photonics co-packaged optics switches, designed to connect millions of GPUs across AI factories while cutting laser count 4x and improving power efficiency roughly 3.5x versus traditional optical networking.

Jan 2026

Quantum-X Photonics switches (144 ports at 800Gbps, liquid-cooled) began reaching customers in early 2026, with Spectrum-X Photonics Ethernet variants following from infrastructure partners the same year — positioning co-packaged optics as an increasingly standard requirement for next-generation AI data centers.

Now

NVIDIA develops co-packaged optics switches, including Quantum-X Photonics and Spectrum-X Photonics, for next-generation AI cluster networking.

Dec 2025

Reports indicated the deal had been quietly scaled back to roughly $30 billion, amid Wall Street scrutiny of "circular" arrangements where NVIDIA invests in a customer that uses the funds to buy NVIDIA's own chips.

Feb 2026

Roughly $30 billion in NVIDIA investment was finalized as part of OpenAI's ~$110 billion funding round, though reporting noted that five months after the original announcement, no binding contract had been signed and no funds had changed hands on the broader $100 billion framework.

Now

NVIDIA committed $30 billion as a lead investor in OpenAI's 2026 funding round, alongside its role as OpenAI's primary GPU supplier.

Nov 2025 · Up to $10B investment + Microsoft strategic partnership

NVIDIA and Microsoft jointly announced strategic partnerships with Anthropic, with NVIDIA committing up to $10 billion and Microsoft up to $5 billion — a notable shift for NVIDIA, which had no direct equity relationship with Anthropic before this point despite Anthropic's Claude models running heavily on NVIDIA hardware via its cloud partners.

Mar 2026 · Investment finalized in the same week as the $30B OpenAI deal

NVIDIA finalized its roughly $10 billion stake in Anthropic in the same week it finalized its $30 billion investment in OpenAI, with CEO Jensen Huang saying shortly after that both were likely to be NVIDIA's last major equity checks into either lab, as both were expected to go public later in 2026.

Now

NVIDIA committed up to $10 billion to invest in Anthropic alongside a strategic partnership with Microsoft, and will help optimize Claude models for performance, efficiency and total cost of ownership.

Jan 2024 · Meta becomes NVIDIA's largest single H100 buyer

Meta became the largest individual buyer of NVIDIA's H100 GPUs, even as it rolled out an improved MTIA v2 ("MTIA 300") with roughly 3x the performance-per-watt of v1 for its ranking and recommendation workloads.

Feb 2026 · Meta buys "millions more" NVIDIA chips despite custom silicon push

Meta agreed to buy millions more AI chips from NVIDIA (spanning Blackwell and the newly announced Vera Rubin platform) even while announcing four new generations of custom MTIA silicon (300/400/450/500) weeks later — underscoring that custom chips are supplementing, not replacing, NVIDIA as Meta's primary AI compute supplier.

Now

NVIDIA supplies millions of AI chips for Meta's data center build-out under an expanded 2026 partnership, a deal analysts describe as certainly in the tens of billions of dollars.

Jul 2025

Licensed H20 sales to China were approved again after the earlier freeze, following a policy reversal easing the country-tiered restriction approach.

Dec 2025

The US administration announced a one-year waiver of export restrictions on NVIDIA's H200 chips, giving Chinese firms legal access to NVIDIA's Hopper chip line.

Now

NVIDIA's chip sales to China are regulated by US export controls, which determine which chip generations can be sold and under what licensing conditions.

Jan 1993

Jensen Huang co-founded NVIDIA and has served as CEO continuously since, guiding the company from a graphics-chip startup to the central supplier of AI accelerated-computing infrastructure.

Now

NVIDIA was co-founded by Jensen Huang, who has served as its CEO throughout the company's history.

Jan 2024

xAI's Colossus in Memphis scaled from roughly 100,000 to 200,000 NVIDIA H100-class GPUs, one of the fastest large-scale GPU cluster buildouts to date.

2025–2026

xAI acquired land for a Colossus 2 expansion and is targeting roughly 555,000 NVIDIA GPUs and about 2GW of power capacity as it scales toward frontier-model training.

Now

NVIDIA supplies the GPUs powering xAI's Colossus supercomputer clusters, among the largest known concentrations of NVIDIA hardware built for frontier AI training.

Jan 2026

The UAE received its first shipment of advanced NVIDIA chips for the Stargate UAE campus developed with G42, following a U.S. government export-license framework enabling the deal.

Now

NVIDIA supplies chips for the UAE's Stargate campus, developed by G42, with the UAE receiving its first NVIDIA chip shipment in 2026.

2025–2026

NVIDIA agreed to supply GB300 Blackwell Ultra chips to HUMAIN, Saudi Arabia's PIF-backed AI company, as part of the kingdom's push to build sovereign national AI compute capacity.

Now

NVIDIA supplies GB300 chips to Saudi Arabia's HUMAIN, backed by the Public Investment Fund, to build out national AI compute capacity.

Jan 2026 · HBM4 volume/pricing contracts finalized alongside Micron and Samsung

SK hynix, alongside Micron and (with more limited volume) Samsung after its delayed HBM3E qualification finally passed in September 2025, moved into finalized HBM4 volume and pricing contracts with NVIDIA for the Vera Rubin platform — SK hynix retaining the largest share of the three qualified suppliers.

Jul 2026 · SK Group and NVIDIA sign $500B+ letters of intent spanning AI factories and HBM4

SK Group and NVIDIA signed letters of intent for a partnership valued at more than $500 billion, under which SK Telecom will build a 2-gigawatt Korean AI data center on NVIDIA's Vera Rubin platform (first phase targeted for 2027) while SK hynix deepens co-development of HBM4 memory with NVIDIA, aimed at easing the memory supply constraints that have become a binding bottleneck on global AI system scaling.

Now

SK hynix supplies advanced memory products used in NVIDIA artificial intelligence computing platforms.

2024–2026

NVIDIA GPUs remained the dominant compute platform behind hyperscaler AI data-center buildouts as combined big-four capex rose from roughly $226B (2024) to $410B (2025) to a projected $725B (2026).

Now

Many AI data-center deployments depend on NVIDIA accelerated computing platforms.

Sep 2025

NVIDIA and OpenAI announced a $100 billion strategic partnership tied to 10 gigawatts of NVIDIA systems deployment, positioning NVIDIA as both OpenAI's primary compute supplier and a lead investor in the same buildout.

Feb 2026

The investment side of the arrangement was finalized at roughly $30 billion within OpenAI's ~$110 billion funding round, while the GPU-supply relationship continued independent of the broader $100 billion framework still being finalized.

Now

OpenAI depends on massive NVIDIA GPU purchases for model training and inference, with NVIDIA committing $30 billion as a lead investor in OpenAI's 2026 funding round.

2024–2026

Oracle Cloud Infrastructure scaled large NVIDIA GPU clusters (including Blackwell-based systems) to support major AI training workloads, including its role in the Stargate data-center initiative.

Now

Oracle depends on NVIDIA GPU supply for its AI training infrastructure, including data centers built around large NVIDIA GPU clusters.

Jan 2024 · Three successor chips planned (Braga, Braga-R, Clea)

Microsoft began developing three successor generations to Maia 100 (codenamed Braga, Braga-R, and Clea, planned for 2025-2027) even as it remained NVIDIA's largest customer by revenue, spending billions annually on NVIDIA chips for Azure.

Jan 2026 · Maia 200 ships, explicitly targeting reduced NVIDIA reliance

Microsoft shipped its second-generation Maia 200 chip (delayed from an original 2025 target) to its Iowa data center, explicitly framed as reducing reliance on NVIDIA chips for inference workloads — while continuing to depend on NVIDIA GPUs as its primary AI training infrastructure.

Now

Microsoft purchases large volumes of NVIDIA GPUs to build out Azure's AI infrastructure.

Jan 2024 · Dual-track spending: NVIDIA GPUs and Trainium2 scale together

Amazon deployed large volumes of NVIDIA Hopper GPUs while simultaneously investing billions in Trainium2 clusters, as combined hyperscaler AI capex roughly tripled industry-wide from about $226 billion in 2024 toward $410 billion in 2025 — Amazon's NVIDIA and custom-silicon spending both grew rather than one displacing the other.

Oct 2025 · Trainium sales commitments reach $225B; ASIC share still growing fast

Amazon's Trainium sales commitments reached roughly $225 billion (including from OpenAI and Anthropic), with AWS's ASIC shipments expected to more than double in 2025 — the fastest growth among major cloud providers — while Amazon continued deploying NVIDIA Blackwell GPUs at scale for the training workloads Trainium doesn't yet fully cover.

Now

Amazon purchases NVIDIA GPUs at scale for AWS AI infrastructure alongside its own custom silicon.

2024–2025

Foxconn served as the sole system assembler for NVIDIA's Blackwell GB200 AI servers, making its Mexico and Taiwan assembly capacity a critical bottleneck during the widely reported Blackwell production ramp delays.

Now

Foxconn assembles NVIDIA's Blackwell GB200 AI servers as the sole assembler, making it a critical bottleneck point in the AI data center supply chain.