COMPUTE ATLAS Supercomputer supply-chain graph
1345 systems 511 sites Sourced data

Updated daily

News

What happened, and which part of the graph it happened to. Every item links out to a primary source and links in to the systems, parts and companies it concerns, so a corporate move shows up on the affected company's page, and an export-control action shows up on the machines it constrained.

Every summary here is ours. We record a headline, an attributed link to the source, and a note written for this site. We do not reproduce anyone else's article text: the link is where you read their words. Items proposed by the daily routine are marked unreviewed until someone has read the source and written the note.

2026

  • Facility rep

    Iren targets 900MW data center campus in Wagga Wagga, Australia

    Iren, an existing GPU-cloud operator in this dataset, outlined a 900MW campus proposal near Wagga Wagga targeting a 2029 grid connection and up to AU$10bn of investment -- its second major Australian campus disclosed this year, and one the company says sits outside its already-announced 5GW secured-power portfolio, so this is pipeline expansion rather than confirmed new capacity.

    Data Center Dynamics → IREN
  • Facility rep

    Sberbank data center in Russia struck by Ukrainian military

    Sberbank confirmed a drone strike on its Balakovo data center in Saratov region, the latest in a wave of strikes on Russian operators this month that has already hit three of Yandex's facilities; Sberbank said services were unaffected but gave no damage detail, adding another row to the physical-security risk pattern already carried for Russian compute operators rather than changing any system's recorded status.

    Data Center Dynamics → Sberbank
  • Procurement rep

    Fermi and AMD sign credit support deal for build-out of Project Matador campus in Amarillo, Texas

    AMD signed a credit-support agreement guaranteeing financing for the first 222MW phase of Fermi America's Project Matador campus, plus a framework for up to 650MW of future capacity, and confirmed thousands of its Instinct GPUs will run there under TensorWave's existing lease -- a chip-vendor guarantee underwriting a buildout whose original anchor tenant had previously withdrawn.

    Data Center Dynamics → Fermi America AMD
  • Corporate rep

    OpenAI again shakes up data center leadership, with xAI exec Uday Ruddarraju to lead and Sachin Katti to leave

    OpenAI moved its compute and infrastructure organisation under a former xAI infrastructure lead, folding commercial strategy and supply-chain decisions into engineering. It is a change in who steers one of the largest GPU buyers' build-out, relevant to how Stargate-era capacity commitments to Nvidia, AMD and cloud partners are sequenced.

    Data Center Dynamics → OpenAI
  • Facility rep

    NTT to build $6.3bn data center for Anthropic in Japan - report

    Unnamed-source reporting (Nikkei Asia, relayed here) says NTT Data is close to signing on to build Anthropic's first dedicated Japanese facility, a $6.3bn, up-to-500MW site in Osaka with partial operations targeted for 2028. Neither company confirmed it, but it would mark Anthropic's first Asian facility and another entry in its already roughly 15GW book of compute commitments.

    Data Center Dynamics → Anthropic
  • Facility rep

    Oracle trucks in gas to support operations at Utah and Texas data centers - report

    Bloomberg's unnamed-source reporting describes Oracle trucking in compressed natural gas to keep its Salt Lake City site and an OpenAI-linked Shackelford County, Texas campus running while pipeline connections lag, and says it is weighing the same workaround for the stalled 2.5GW Project Jupiter site in New Mexico after regulators blocked a pipeline extension there. A concrete case of on-site power delivery, not chip supply, becoming the binding constraint on bringing new AI capacity online.

    Data Center Dynamics → Oracle
  • Corporate rep

    d-Matrix Announces Squadron Compute Clusters for AI Inference at Data Center Scale

    d-Matrix, built around its own memory-centric Corsair XPU chip, is now shipping full air-cooled, multi-rack "Squadron" inference clusters (up to 1,024 XPUs) built with Supermicro, Arista and Broadcom, already running at its own Arizona site and now shipping to outside customers -- a new merchant-silicon entrant at the inference-accelerator layer this graph tracks, pitched for heterogeneous GPU-plus-XPU deployments rather than as a drop-in GPU replacement.

    HPCwire → Supermicro Arista Networks Broadcom
  • Facility rep

    Third strike on Yandex data center in Russia

    A third drone strike in a week took down Yandex's Vladimir facility, pausing dozens of cloud services; it follows the Sasovo strike already in this dataset, which outside reporting says holds two of Yandex's three machine-learning supercomputers, so this is another entry in the physical-security risk this graph's Yandex facility rows carry, not a change to any system's recorded status or location.

    Data Center Dynamics → Yandex
  • Corporate rep

    Microsoft Boosts HBM Bandwidth by 2.5x with Maia 300 Accelerator

    Microsoft's own technical brief lays out Maia 300 succeeding Maia 200: a shift from a monolithic 3nm die to a chiplet design with NUMA memory access and a two-tier scale-up network that supports superpods of over 12,000 accelerators, roughly doubling Maia 200's largest cluster size. It's a second generation of in-house Azure inference silicon, relevant to how much of Microsoft's AI fleet runs on custom ASICs rather than merchant GPUs.

    HPCwire → Microsoft
  • Corporate rep

    Sponsored: Wärtsilä, Schneider Electric, and Stanley Consultants launch coordinated approach for faster US data center power delivery

    Three power-delivery vendors announced a bundled 'generator-to-chip' offering -- onsite generation, electrical distribution and engineering under one coordinated contract -- aimed at US AI data centers; no project, site or capacity is named, but it's a vendor-side signal that grid-interconnection delays are pushing power suppliers toward turnkey, pre-integrated delivery rather than developers assembling generation and electrical scope separately.

    Data Center Dynamics → Schneider Electric
  • Corporate rep

    Report: Nvidia plans to invest in inference chip startup d-Matrix

    Unnamed-source reporting that Nvidia is planning an equity investment in d-Matrix, a month after d-Matrix adopted Nvidia's NVLink Fusion to link its own inference chips with Nvidia hardware (already noted in this dataset). An investment would tie that merchant-silicon partner more closely to Nvidia's own roadmap rather than keeping it independent; single-sourced via The Information, so treated as reported, not confirmed by either company.

    Data Center Dynamics → NVIDIA
  • Facility rep

    Second Yandex data center hit in drone strikes in Russia

    A second drone strike in 48 hours hit a Yandex data center, this one in Kaluga, after an earlier strike disabled the Sasovo site that outside reporting (Reuters, cited in the article) says holds two of Yandex's three machine-learning supercomputers in this graph. Our own rows for those systems deliberately leave the physical site blank for lack of a primary source, so this is recorded as a facility risk to watch rather than a change to any system's location or status.

    Data Center Dynamics → Yandex
  • Facility rep

    Anthropic-linked $366m data center project filed for in Cedar Creek, Texas

    A Texas filing puts a $366M, roughly 490MW data center building in Bastrop County on the schedule, with Anthropic named as the likely tenant even though none of the parties involved will confirm it. The power plan leans on a separate 710MW onsite gas turbine plant rather than the regional grid, arriving just as Texas regulators have frozen new data center grid-interconnection permits pending an ERCOT capacity audit, and lands the project next to EdgeConneX's existing multi-building campus in the same county.

    Data Center Dynamics → Anthropic
  • Facility rep

    Sesterce to invest $10bn in 600MW AI data center campus in Jämsä, Finland

    Sesterce, already tracked here for its Valence, France buildout, is now planning a second-country campus on a former paper-mill site in Finland with long-term ambitions past 1GW, which tests whether a mid-tier AI-cloud entrant can keep scaling its own sites rather than just lease capacity from larger operators.

    Data Center Dynamics → Sesterce
  • Facility rep

    Yandex data center hit in drone attack in Russia

    A drone strike disabled Yandex's largest Russian data center and knocked out an entire cloud availability zone, the first time the war has taken down a major facility inside Russia rather than Ukraine, which is a reminder that the graph's facility rows carry physical-security exposure that capacity figures alone don't capture.

    Data Center Dynamics → Yandex
  • Corporate rep

    HPE Launches Gen13 ProLiant Servers for HPC and AI Workloads

    HPE's Gen13 ProLiant line pairs AMD's next-gen EPYC and upcoming "Venice" CPUs with Nvidia BlueField DPUs and ConnectX NICs across both legacy 19-inch and OCP's new 21-inch rack formats, a concrete look at the hardware generation vendors will draw on for the next wave of HPC/AI deployments -- notable given HPE's track record as sole compute, fabric and storage supplier on both Frontier and El Capitan.

    HPCwire → Hewlett Packard Enterprise
  • Facility rep

    Applied Digital to develop 1GW data center campus in Finland

    Applied Digital's first development outside the US is only a power-access agreement with no site or tenant named yet, so the 1GW figure is pipeline capacity, not contracted load; it shows the same Dakota/Louisiana/Alabama AI-campus playbook now reaching for Nordic power and climate advantages.

    Data Center Dynamics → Applied Digital
  • Corporate rep

    ByteDance takes up one-fifth of China's data center capacity

    A SemiAnalysis count of China's data-center market puts a number on how concentrated hyperscaler leasing is: ByteDance alone holds roughly a fifth of a 24GW-plus market, almost entirely leased rather than self-built, unlike Huawei's largely self-build model. A structural data point on who actually controls AI compute capacity in China versus who merely operates it.

    Data Center Dynamics → ByteDance
  • Milestone rep

    Germany’s Karlsruhe Institute of Technology unveils 33 petaflops HoreKa 2 supercomputer

    KIT's HoreKa 2 goes operational as the successor to the HoreKa system already tracked in this graph, roughly doubling peak performance to 33 petaflops on a Lenovo-built hybrid of AMD Epyc Turin CPUs and Nvidia GB200 GPUs. A concrete data point for how NHR centers refresh hardware generations, and reason to update HoreKa's own system entry with the newer configuration.

    Data Center Dynamics → HoreKa Karlsruhe Institute of Technology
  • Facility rep

    Bull Expands Supercomputer Manufacturing with €80M Angers Redevelopment

    Angers is the one European plant assembling Eviden's BullSequana systems, including the Alice Recoque and LUMI-AI builds already tracked here, so this €80M capacity doubling (due 2028) is a concrete signal on whether those deliveries stay on schedule. It also surfaces a new Foxconn partnership to build NVIDIA Vera Rubin NVL72 AI servers in Europe, relevant to the EU domestic-substitution question raised elsewhere in this dataset.

    HPCwire → Bull Atos Eviden
  • Facility rep

    Crusoe files for $4.8bn data center campus in Jayton, Texas

    Two more spur buildings of Crusoe's Project Hyper campus, filed in Jayton near its existing Childress and Abilene Texas sites, bring Crusoe's disclosed spend across the Jayton and Childress filings to roughly $16.8bn across five buildings. Useful for tracking Crusoe's pace and footprint as a developer/landlord, distinct from the GPU-cloud deals under its own brand already in the graph.

    Data Center Dynamics → Crusoe
  • Facility rep

    AI data center planned for Karlsruhe Institute of Technology, Germany

    KIT already operates HoreKa, the national NHR supercomputer tracked in this graph; this is a separate, smaller dedicated AI facility at its Campus North site, metered initially at 2MW with headroom to 10MW, built next to a CHP plant for waste-heat reuse into district heating. Facility-layer data point on a research institution's compute footprint, distinct in scale from the hyperscaler builds already logged here.

    Data Center Dynamics → Karlsruhe Institute of Technology
  • Corporate rep

    Optical interconnect startup CScale emerges from stealth following investment from Nvidia and Intel

    CScale is a new entrant in the optical-interconnect layer, backed by strategic investment from both Nvidia and Intel inside a $145M round; worth tracking because gigawatt-class AI campuses are pushing rack-to-rack interconnect toward optics, and money from two accelerator vendors signals where that transition may land.

    Data Center Dynamics → NVIDIA Intel
  • Corporate rep

    CoreWeave makes Nvidia Vera Rubin NVL72 available, Cognition is early customer

    CoreWeave is now offering Nvidia's Vera Rubin NVL72 rack on its cloud, with Cognition as the first named customer; extends the existing CoreWeave-Nvidia relationship already in the graph and gives an early, vendor-reported throughput data point for Rubin ahead of the systems that will eventually carry it as a component.

    Data Center Dynamics → NVIDIA CoreWeave NVIDIA Vera Rubin
  • Facility rep

    Google named as customer for Crusoe's data center campus in Amarillo, Texas

    Crusoe confirmed Google as the anchor customer for its multi-building Amarillo campus, sited next to and powered by Crusoe's own wind farms; a useful data point on Crusoe's role as developer/landlord rather than compute operator, distinct from the GPU-cloud deals under Crusoe's own brand already tracked here.

    Data Center Dynamics → Crusoe
  • Procurement rep

    Gimlet Cloud partners with Cerebras to deploy 100MW of wafer-scale compute

    A newly well-funded neocloud is committing to 100MW of Cerebras wafer-scale inference capacity and will be a launch partner for Cerebras' next CS-4 chip; a rare procurement data point for an accelerator vendor outside the Nvidia-dominated share this graph otherwise tracks.

    Data Center Dynamics → Cerebras
  • Procurement rep

    HPE Secures Its 1st AMD Helios Order in $1.2B Deal with Vultr

    HPE's first order for AMD's Helios rack platform: a $1.2B deal to supply Vultr with racks of Instinct MI455X GPUs, EPYC "Venice" CPUs and HPE-Juniper scale-up Ethernet fabric, with no unit count disclosed. Worth flagging that "Helios" here is AMD's rack-scale product name, not a single deployed system, and is unrelated to the two existing Helios entries already in this graph (Cyfronet's Poland machine and the decommissioned Japan fusion-research system) beyond the shared name.

    HPCwire → AMD Vultr Hewlett Packard Enterprise
  • Facility rep

    Amazon to acquire two data centers in Hong Kong

    Amazon is converting two data-center buildings it already leases in Hong Kong -- iTech Tower 3.1 and 3.2 -- into owned assets, paying landlord Grand Ming a minimum of about US$278m (up to US$312m) as Grand Ming sells to deleverage. It's a small but checkable data point on a hyperscaler shifting from leasing to owning facility real estate in a supply-constrained market, distinct from new capacity coming online.

    Data Center Dynamics → Amazon
  • Corporate rep

    Schneider Electric and Equinix trial software-defined switchgear for data centers

    Equinix is piloting Schneider Electric's new software-defined medium-voltage switchgear at an undisclosed site, a design Schneider says cuts control wiring, terminals and copper sharply versus engineered-to-order gear. It's a concrete move on the power-delivery-hardware side of the supply chain this graph already tracks alongside NVIDIA's DSX Ready program, aimed at speeding up how fast electrical infrastructure for AI-scale data centers can be built.

    Data Center Dynamics → Schneider Electric
  • Facility rep

    Lambda plans data center in Oklahoma

    Lambda, an NVIDIA-GPU cloud provider whose known customers include Microsoft and Anthropic, is developing a new data center outside Tulsa in Oklahoma's MidAmerica Industrial Park (also home to a Google campus), with closed-loop cooling, grid-reactive load design, and about $500m in committed local tax payments over ten years. It's another facility-layer capacity addition for a neocloud already in this graph, notable for structuring the deal so Lambda -- not ratepayers -- covers the site's power costs.

    Data Center Dynamics → Lambda
  • Facility rep

    Vertiv expands data center kit manufacturing capacity in Slovakia

    Vertiv is adding roughly 22,000 sq m at its Slovak campus over the next two years for power, switchgear and liquid-cooling equipment, which is the kind of non-accelerator capacity the graph tracks as a possible bottleneck for high-density AI and HPC racks. Regional trade-press coverage of a company announcement, with no investment figure given, so recorded as reported.

    TechXMedia → Vertiv
  • Corporate rep

    Vertiv to acquire data center fluid management firm King Environmental Services

    Vertiv's second fluid-management acquisition in under a year, after 2025's PurgeRite deal in North America, extends its cooling-services reach into EMEA -- a sign that liquid-cooling logistics (fluid prep, flushing, load-testing) is consolidating as its own sub-market alongside the hardware itself as GPU rack density rises.

    Data Center Dynamics → Vertiv
  • Corporate rep

    AI cloud and data center firm Nscale raises $3.36bn ahead of its IPO

    Nscale's $3.36bn pre-IPO round includes a $1bn tranche from NVIDIA itself, adding an equity stake to the vendor-customer relationship already in the graph and funding the roughly 461,000-GPU buildout -- mostly NVIDIA Vera Rubin NVL72 -- including capacity contracted by Microsoft and Anthropic at the new Monarch Compute Campus in West Virginia.

    Data Center Dynamics → NVIDIA Nscale
  • Corporate rep

    DensityAI: Chip startup founded by former Tesla Dojo execs valued at $10bn – report

    Former leads of Tesla's Dojo chip program, shut down in August 2025, have founded DensityAI, reportedly raising a $10bn-valuation round with a performance-contingent purchase agreement from AWS for a new accelerator built around 3D-stacked DRAM -- a concrete example of an abandoned in-house chip effort seeding an independent merchant-silicon entrant. Sourced to a single outlet (The Information) with Amazon, Andreessen Horowitz and DensityAI all declining to confirm the details, so treat as reported rather than verified.

    Data Center Dynamics → Tesla Dojo Tesla
  • Corporate rep

    NVIDIA Launches DSX Ready to Qualify Power and Cooling Products for AI Factories

    NVIDIA's new DSX Ready program formalizes third-party qualification of power and cooling hardware against its AI-factory reference design, launching with battery storage from Hitachi Energy, LG Energy Solution and Tesla and cooling-distribution units from LG Electronics, LiquidStack and Vertiv -- a sign that the compute supply chain's constraint is shifting from the accelerators themselves toward the power and thermal infrastructure needed to run them at rack scale. Single-sourced to NVIDIA's own announcement (republished by HPCwire), so recorded as reported rather than verified.

    HPCwire → NVIDIA Tesla Vertiv
  • Research rep

    LightSolver and HLRS Explore Photonic Acceleration for HPC Simulation Workloads

    An HLRS/LightSolver paper reports emulator-based projections that an experimental all-optical processor could sharply speed up sparse linear-solver workloads versus GPUs, but the results are simulated rather than measured on physical hardware and the LPU has not been built at HPC scale. It's an early research signal on whether photonic accelerators could eventually join CPUs and GPUs in production systems, not evidence of a deployment.

    HPCwire → HLRS
  • Corporate rep

    QuEra partners with HPE to integrate quantum computer within Cray-powered HPC environments

    HPE is opening its Cray-based HPC systems to QuEra's neutral-atom quantum hardware, giving HPC centers and national labs an on-premises path to quantum-classical integration alongside QuEra's existing AWS Braket cloud access. It matters as a structural signal that a major classical-HPC integrator is formalizing quantum-accelerator integration rather than leaving it to cloud-only access.

    Data Center Dynamics → Cray Hewlett Packard Enterprise
  • Facility rep

    Oracle issues force majeure notice to Blue Owl following series of setbacks at Project Jupiter data center campus

    Oracle's force-majeure notice to landlord Blue Owl is a bid to delay lease payments rather than exit its commitment to the 2.5GW Project Jupiter campus in New Mexico, after state regulators rejected the gas-pipeline extension the site needs -- a concrete case of permitting and power delivery, not chip supply, being the binding constraint on a gigawatt-scale AI buildout. Note: this campus is unrelated to the JUPITER exascale system at Julich already in this dataset; the shared name is coincidental.

    Data Center Dynamics → Oracle
  • Milestone rep

    SiPearl Delivers 1st Rhea1 CPU Samples to Bull for JUPITER Supercomputer

    SiPearl shipped the first Rhea1 samples to Bull for integration onto the BullSequana XH3000 platform that forms JUPITER's CPU cluster module, the first delivery of a European-designed processor into a European-built exascale system and SiPearl's first revenue-bearing order. It's a concrete step on the domestic-substitution question this graph already flags for Rhea1: a European host CPU moving from lab benchmarking into an actual EuroHPC machine rather than staying a roadmap item.

    HPCwire → JUPITER Bull SiPearl SiPearl Rhea1
  • Procurement rep

    Prysmian, Rio Tinto to supply Amazon data center in Ohio with low-carbon aluminum

    A materials-supply link in Amazon's Ohio buildout: Prysmian will manufacture electrical cabling from Rio Tinto and Alcoa's Elysis low-carbon aluminum for an unnamed AWS data center, the second Amazon-Rio Tinto low-carbon materials deal after a copper agreement in January. It's a small but checkable data point on hyperscalers extending decarbonization commitments into power-delivery hardware, not just chips.

    Data Center Dynamics → Amazon
  • Corporate rep

    NVIDIA Launches DSX Ready to Qualify Power and Cooling Products for AI Factories

    NVIDIA's new DSX Ready program formally qualifies third-party power and cooling hardware against its AI-factory reference design, starting with battery storage (Hitachi Energy, LG Energy Solution, Tesla) and liquid-cooling CDUs (LG Electronics, LiquidStack, Vertiv). It's a structural move: NVIDIA is now gatekeeping which power/cooling vendors count as compatible with its rack designs, which shapes who wins facility-layer contracts downstream of the GPU purchase.

    NVIDIA Blog → NVIDIA Vertiv Tesla
  • Policy rep

    Google, Nvidia, and Emerald AI found the AI Energy Management Alliance to support demand response capabilities within the data center sector

    Google, Nvidia and Emerald AI founded an alliance, joined by Anthropic, RWE and several grid operators, to push AI data centers toward dispatchable, demand-responsive grid participation instead of fixed capacity draw. That's directly relevant to this graph's recurring finding that power delivery, not chip supply, is now the binding constraint on GPU rack density.

    Data Center Dynamics → NVIDIA
  • Procurement rep

    Bull Beats Out HPE For Next-Gen Lumi AI Supercomputer

    EuroHPC and CSC Finland picked Bull over incumbent integrator HPE to build the LUMI-AI upgrade partition, with delivery slated for the second half of 2027. Worth tracking on the graph because it's a rare case of an existing HPE/Slingshot machine losing its re-up bid to a European vendor running its own BXI interconnect, which fits EuroHPC's stated preference for buying European where it can.

    The Next Platform → LUMI Bull Hewlett Packard Enterprise
  • Corporate rep

    Emerald AI, Google and NVIDIA Launch Alliance to Advance Flexible AI Data Centers

    NVIDIA, Google and startup Emerald AI founded the AI Energy Management Alliance to set technology-neutral standards letting AI data centers shift load, discharge storage or curtail draw on signal from the grid, instead of connecting as fixed always-on demand. It matters here because it is the two largest buyers of AI accelerators publicly treating grid power, not chip supply, as the binding constraint on how fast their own compute buildouts can scale, and because flexibility commitments like these are what utilities are starting to condition faster interconnection on.

    NVIDIA Blog → NVIDIA Alphabet
  • Policy rep

    IBM’s Anderon Finalizes $1B CHIPS Award for Quantum Wafer Manufacturing

    Anderon, IBM's quantum-foundry subsidiary, finalized a $1B CHIPS and Science Act award, matched by a further $1B from IBM, for its 300mm quantum-wafer fab in Albany, which the company says is already producing wafers commercially rather than at research scale. Worth tracking as a rare case of CHIPS money going toward quantum hardware capacity specifically rather than classical logic or memory fabs, which is a different supply chain than the accelerator and interconnect chains this site otherwise follows.

    HPCwire → IBM
  • Facility rep

    Microsoft files to build data center campus in Prince William County, Virginia

    Microsoft filed federal and state permits for a two-building data center campus on 123 acres it bought for $465.5M in 2024, directly bordering an Amazon campus and near a Google-linked site in the same Gainesville corridor. It's another data point on how concentrated the current AI buildout is in a handful of Northern Virginia counties, with the largest hyperscalers now building adjacent to rather than in competition for the scarce sites there.

    Data Center Dynamics → Microsoft
  • Corporate rep

    Schneider Electric debuts 2.5MW power modules for AI data centers

    Schneider Electric's new prefabricated power modules, sized up to 2.5MW (its largest yet) and built to a reference design co-developed with NVIDIA for the GB300 NVL72 rack, are meant to cut the time to stand up hyperscale power infrastructure from years to months. It's a facility-layer signal that power delivery itself is becoming a standardized, vendor-packaged product as AI data centers scale, which matters to the graph's facility side as much as any accelerator announcement does.

    Data Center Dynamics → Schneider Electric NVIDIA
  • Facility rep

    Meta to expand Eagle Mountain data center campus in Utah

    Meta's fourth expansion of its Eagle Mountain, Utah campus pushes its investment there past $3 billion, adding self-built capacity outside the HPL-measured world this site otherwise tracks -- the same disclosed-versus-undisclosed distinction that motivates the shadow-list page. The new buildings will use closed-loop liquid cooling with dry cooling, a facility detail worth watching as more hyperscaler capacity shifts away from air.

    Data Center Dynamics → Meta Platforms
  • Corporate rep

    SiFive and AMD Collaborate to Optimize AMD ROCm on RISC-V Datacenter Servers

    SiFive and AMD showed AMD's ROCm software stack running an inference workload on a server built around SiFive's own RISC-V cores as head nodes, paired with AMD Radeon AI PRO R9700 GPUs, at a trade-show demo rather than as a shipping product. It's a small but concrete signal that RISC-V is being positioned as a third CPU option inside GPU-accelerated nodes alongside x86 and Arm, which would matter to the graph's CPU-vendor mix if it moves past demo hardware.

    HPCwire → AMD
  • Corporate rep

    From Megawatts to Tokens: How NVIDIA Maximizes AI Factory Production

    Nvidia's new DSX platform cites a Lambda cloud benchmark of 24% more token throughput from the same power budget on HGX B200 clusters, and names a 96MW Vera Rubin NVL72 facility in Manassas, VA as its first dedicated grid-flexibility deployment. Concrete power-per-token figures worth tracking, though all of them come from Nvidia's own writeup and a partner quote rather than an independent measurement.

    NVIDIA Blog → NVIDIA
  • Procurement rep

    Bull awarded €31m contract with AEMET to deliver infrastructure for weather forecasting supercomputer

    A four-year, €31m contract for AEMET's Madrid-hosted forecasting supercomputer, awarded to Bull only months after the French state completed its buyout of Bull from Atos — a data point on Bull's post-acquisition order book alongside its Mimer and LUMI-AI wins earlier this year.

    Data Center Dynamics → Bull Atos
  • Corporate rep

    Microsoft targets 38GW of data center capacity in 2032 - report

    Microsoft's reported target of roughly tripling to 38GW of capacity by 2032, from about 12GW now, is a forward plan sourced to Bloomberg's unnamed sources rather than a confirmed figure, so it belongs here as stated intent rather than anything checkable against installed hardware. The structural point for the graph is how much of that growth is meant to come from leasing neocloud capacity (CoreWeave, Nscale, Lambda, Iren, Nebius were named) rather than built and owned outright, which is exactly the operator/owner distinction entity resolution here has to keep straight rather than collapse into one Microsoft footprint.

    Data Center Dynamics → Microsoft
  • Corporate rep

    Google and Blackstone's TPU venture launches as Crux AI with a stated 2GW capacity target

    TPUs sit in this graph as parts inside Google's own infrastructure; a separately capitalised operator selling TPU capacity puts the same accelerator on the far side of a company boundary, which is exactly the case entity resolution here has to get right rather than collapse into its parent. The capacity numbers are the company's own forward plans and an executive hire, not built floor space, so they establish intent and nothing more.

    Data Center Dynamics → Alphabet
  • Corporate rep

    NVIDIA opens NVLink Fusion to d-Matrix inference silicon for rack-scale deployment

    NVLink enters the parts layer as the fabric binding NVIDIA's own accelerators inside a rack; licensing it to a third party's inference silicon would move it toward a merchant interconnect competing with the InfiniBand and Ethernet rows beside it. This is an announced intention around an unreleased part, sourced to the vendor's own blog, so no system in the graph attaches to it.

    NVIDIA Blog → NVIDIA
  • Procurement rep

    IQM Sets Three-Stage Upgrade Path for LUMI-IQ Quantum System

    This is a procurement commitment, not a delivered system: CSC's LUMI site in Kajaani gains an IQM quantum computer on a three-stage path, from a 2027 Halocene H4 install to a 2029 Halocene H5 upgrade, sourced entirely to the vendor's own press release. It matters to the graph because it ties a quantum system to a site LUMI already occupies here, well before there is any hardware to add as a component row.

    HPCwire → LUMI
  • Corporate rep

    Broadcom Rides Rocketing Trend For Custom AI Accelerators

    Broadcom's reported AI revenue and forward guidance point to custom accelerator programs (Google, Anthropic, OpenAI, Meta) scaling into multi-gigawatt annual deployments, which matters for the graph because it marks a structural shift of hyperscaler capacity planning away from merchant GPUs and toward vendor-specific silicon built through Broadcom's packaging and TSMC manufacturing, with HBM supply as the binding constraint on how fast that shift can go.

    The Next Platform → Broadcom
  • Research rep

    Panmnesia and Meta propose extending cache-coherent CXL beyond the rack to datacenter scale

    Every interconnect row in this dataset describes a fabric that either stops at the rack or treats cross-rack traffic as a network hop; the proposal here is to hold one cache-coherent domain across a whole facility, which would turn that boundary from a given into a design choice. It is an invited review article with components claimed as validated silicon but no deployed system behind it, so nothing in it belongs in the parts layer yet.

    HPCwire → Meta Platforms
  • Research rep

    Intel Foundry reports High-NA EUV running in volume production on select 18A layers

    The accelerator layer of this dataset is fabricated almost entirely at one foundry, so whether a second leading-edge source reaches production is a question sitting upstream of every part row here. The wafers described are a client CPU rather than a datacenter part, which is the limit of what this establishes.

    Intel Newsroom → Intel
  • Corporate rep

    HPE says component price increases are passing through to buyers without costing it orders

    HPE is the integrator of record on both Frontier and LUMI, so its input costs sit directly underneath two of the largest systems in the graph. Rising prices at the DRAM, accelerator and fabric layers that do not suppress orders is a supply-side constraint showing up in the parts layer rather than a demand story.

    The Next Platform → Hewlett Packard Enterprise
  • Corporate rep

    Dell reports a $95bn AI systems backlog alongside a forecast that AI reaches 75% of datacenter demand by 2030

    Dell is the integrator of record on Frontera and Stampede2 in this graph, so its order book sits directly on top of the same accelerator and fabric supply every other integrator here draws from, and a backlog several times last year's is a demand signal pressing on that shared upstream. The backlog and quarterly sales are reported results; the 2030 demand share, token growth and power figures beside them are the company's own forecast and are not checkable against anything in this dataset.

    The Next Platform → Dell Technologies
  • Policy rep

    Louisiana regulators fast-track review of seven more gas plants for Meta's Hyperion campus

    Ten gas fired units totaling roughly 7.5 gigawatts, up from three approved in August 2025, would power a single Meta data center at a scale that a critic notes exceeds six times New Orleans' peak demand. It is a sharp illustration of the recurring finding elsewhere in this dataset that grid power, not chip supply, has become the binding constraint on gigawatt class AI campuses like Hyperion.

    Union of Concerned Scientists → Meta Hyperion Meta Platforms
  • Corporate rep

    French state completes its acquisition of Bull from Atos

    The state becomes sole owner of Bull's HPC, quantum and AI divisions for up to 404 million euros, nine months after the July 2025 purchase agreement, formalizing Bull as a sovereign supplier rather than a subsidiary of a financially troubled parent. Bull's BullSequana platform underpins JUPITER, so this is a change of ownership sitting directly upstream of an existing exascale system in this dataset, not just a balance sheet event.

    Atos Group → Bull Atos Eviden JUPITER
  • Policy rep

    Commerce Department moves H200 and MI325X exports to China to case-by-case review

    Replaces the blanket license requirement on NVIDIA H200 and AMD MI325X exports with conditional approval tied to US supply, buyer compliance screening and independent US-based testing, following Trump's December 2025 announcement permitting the sales. It is a policy loosening rather than the tightening this dataset's earlier Entity List rows describe, and it applies to parts one tier above the H100/MI300-class accelerators already in the graph.

    US Bureau of Industry and Security → NVIDIA AMD
  • Procurement rep

    Meta signs nuclear power deals with Oklo, Vistra and TerraPower for up to 6.6GW

    Vistra's contribution includes uprates at its existing Perry and Davis-Besse reactors, an operating asset route to new capacity that is faster than the Oklo and TerraPower agreements, which depend on advanced reactor designs not expected online before 2030 and 2032 respectively. Meta's Prometheus campus in Ohio, already in this dataset, is named as one beneficiary, though only a fraction of the contracted capacity is available on any timeline that matches Prometheus's 2026 target.

    Utility Dive → Meta Prometheus Meta Platforms

2025

  • Procurement rep

    Dell and NVIDIA detail Horizon, the NSF's Frontera successor at TACC

    Horizon pairs unreleased NVIDIA Grace Blackwell and Vera parts with Dell's direct liquid cooled IRSS racks and Quantum-X800 InfiniBand, targeted at roughly 10x Frontera's simulation throughput. NSF funding uncertainty flagged earlier in 2025 did not stop the build, and early user access is slated for spring 2026, which is worth checking against whatever actually ships.

    Dell Technologies → Horizon Frontera Dell Technologies NVIDIA
  • Milestone rep

    TOP500's 66th list confirms JUPITER as the first exascale system outside the US

    El Capitan holds number one at a remeasured 1.809 exaflop/s and Frontier and Aurora hold two and three, so all three US exascale systems in this dataset keep their relative order. JUPITER's 1.000 exaflop/s Booster partition becomes the fourth exascale system worldwide and the first in Europe, confirming the milestone this dataset's earlier rows on Bull's JUPITER delivery had been building toward.

    TOP500 → El Capitan Frontier Aurora JUPITER
  • Facility rep

    Microsoft links Fairwater Atlanta and Wisconsin into one "AI superfactory"

    The two Fairwater sites are joined over a dedicated AI WAN fiber backbone so training jobs can span both facilities as one system rather than two separate clusters, a different scaling strategy than the single site gigawatt campuses xAI and Meta are building. Each rack carries up to 72 NVIDIA Blackwell GPUs; Microsoft has not disclosed a total GPU count for either site.

    Microsoft → Microsoft Fairwater (Atlanta) Microsoft Fairwater (Wisconsin) Microsoft NVIDIA
  • Corporate rep

    Google's Ironwood (TPU v7) becomes generally available, compared directly to El Capitan

    Google's own launch materials put a full Ironwood pod at 42.5 exaflops of FP8 compute, described as more than 24x El Capitan's 1.7 exaflops figure, though the two numbers measure different things, dense FP8 inference throughput versus double precision HPL, and are not a like for like comparison. It is nonetheless the most direct public benchmark claim yet tying a hyperscaler's own accelerator generation to a specific system already in this dataset.

    Google → Google TPU7x (Ironwood) pod Alphabet El Capitan
  • Procurement rep

    DOE, NVIDIA and Oracle announce Solstice and Equinox at Argonne

    Solstice, with 100,000 Blackwell GPUs, and Equinox, with 10,000, are billed at a combined 2,200 exaflops of AI performance and sit at the same lab as Aurora, extending Argonne's fleet through a public-private financing model rather than the pure DOE-funded build Aurora was. Equinox is due in 2026; Solstice has no public delivery date yet, so neither belongs in the systems table until hardware lands.

    IT Pro → Solstice Equinox NVIDIA Oracle Argonne National Laboratory
  • Milestone rep

    AWS activates Project Rainier, a nearly 500,000-chip Trainium2 cluster for Anthropic

    Rainier is the largest non-NVIDIA training deployment in this dataset by chip count and the clearest evidence yet that a hyperscaler's own silicon, not merchant GPUs, can anchor a frontier model training run. AWS says Anthropic is headed toward more than 1 million Trainium2 chips by year end, which would make this entry's chip count stale within months rather than years.

    About Amazon → Project Rainier Amazon AWS Trainium2
  • Procurement rep

    ORNL, AMD and HPE announce Discovery and Lux as Frontier's AI-era successors

    Lux runs AMD Instinct MI355X on HPE nodes with Oracle Cloud Infrastructure integration and arrives first, in 2026; Discovery follows in 2028 on MI430X and is billed as beating Frontier on every part of the system, not just peak flops. This keeps ORNL's AMD based lineage going a full hardware generation past the MI300A in El Capitan and the MI250X in Frontier.

    Oak Ridge Leadership Computing Facility → Discovery (ORNL) Lux (ORNL) Frontier AMD Hewlett Packard Enterprise
  • Facility rep

    TSMC's Arizona fab begins volume production of NVIDIA Blackwell wafers

    First Blackwell wafers made on US soil, on a 4nm process two generations behind TSMC's leading edge, with the dies still shipped back to Taiwan for advanced packaging until Arizona packaging capacity comes online around 2027 to 2028. It narrows, without closing, this dataset's single foundry dependency for the accelerator layer behind El Capitan, the Frontier successor systems and most hyperscaler GPU fleets tracked here.

    SiliconANGLE → TSMC NVIDIA
  • Corporate rep

    SK hynix finishes its first HBM4 design, targets NVIDIA's Vera Rubin generation

    12 layer HBM4 at 10Gbps per pin and over 40 percent better power efficiency than HBM3E, delivered after NVIDIA reportedly asked for a six month pull-in versus the original early 2026 timeline. HBM3E already appears in this dataset's parts layer under several 2024-era accelerators; this is the supply chain step that determines when the next memory generation can show up under Vera Rubin systems instead.

    KED Global → SK hynix NVIDIA
  • Research rep

    ETH Zurich and EPFL release Apertus, a fully open LLM trained on Alps

    Apertus used more than 10 million GPU hours on Alps to train 8B and 70B parameter models with open weights, data and training code, a rare case in this dataset of a system's compute being fully accounted for by one named downstream model rather than a mix of undisclosed workloads.

    ETH Zurich → Alps CSCS ETH Zurich
  • Facility rep

    xAI's Colossus 2 reaches 200MW of cooling capacity in six months

    The Memphis and Southaven build reached the milestone in about six months against roughly fifteen months for comparable Oracle, Crusoe and OpenAI projects, and is targeted at over 1GW of GB200 NVL72 capacity. It is a data point on how much faster a single well capitalized operator can now stand up gigawatt class capacity than the shared infrastructure neoclouds this dataset also tracks.

    SemiAnalysis → xAI Colossus 2 xAI NVIDIA
  • Milestone rep

    Isambard-AI phase 2 goes live as the UK's most powerful AI supercomputer

    5,448 GH200 superchips across a purpose built 5MW facility that went from conception to operation in about two years, reported as more than ten times faster than the next UK system at the time. It is the clearest UK entry in this dataset's post-2023 Grace Hopper wave alongside LUMI and Alps.

    NVIDIA Blog → Isambard-AI Phase 2 Hewlett Packard Enterprise NVIDIA University of Bristol
  • Procurement rep

    DOE names Dell and NVIDIA to build NERSC-10, Doudna

    Doudna succeeds Perlmutter at NERSC and is contracted for more than 10x Perlmutter's scientific output on 2 to 3x the power, an efficiency target that puts a specific number on the performance-per-watt gains this dataset already tracks across GPU generations. It ships on unreleased NVIDIA Vera Rubin parts, so nothing in the graph attaches to it yet beyond the contract.

    NVIDIA Blog → Doudna Perlmutter Dell Technologies NVIDIA
  • Milestone rep

    Piz Daint powered off after 13 years, cleared for Alps

    CSCS retired the Cray XC50/XC40 system that had anchored Swiss GPU computing since 2012, closing out the last pre Grace Hopper system in this dataset's Swiss lineage and completing the multi year handover to Alps that began with Alps' 2024 inauguration.

    CSCS → Piz Daint Alps CSCS

2024

2023

2022

2021

  • Policy rep

    Phytium added to the US Entity List

    Extends the substitution question from accelerators to the host CPU: Phytium's Arm server parts are the domestic option underneath the Tianhe line.

    Federal Register → Phytium NUDT

2020

2019

2015

  • Policy ver

    NUDT and the Chinese national supercomputing centres added to the Entity List

    Blocked the planned Xeon Phi upgrade to Tianhe-2. The 2018 rebuild replaced the American accelerators with NUDT's own Matrix-2000 and the machine got faster.

    Federal Register → NUDT Tianhe-2A NUDT Matrix-2000

RSS feed · News is cross-referenced onto system and company pages automatically, so Tianhe-2 carries the export-control items that constrained it.