AI

Tencent has signed a five-year contract to access 100,000 GPUs across multiple Oracle data centers in Southeast Asia, according to the Financial Times, which cited two people familiar with the matter. The deal is valued at approximately $7 billion, with Tencent paying 30 percent upfront. The arrangement allows Tencent to access GPU hardware unavailable in China due to U.S. export controls, with compute intended primarily for training Hunyuan models, followed by inferencing workloads, and eventually for rental via Tencent Cloud. CFO John Lo and President Martin Lau both addressed the company's AI infrastructure spending, with Lau noting that infrastructure could be rented out at cost recovery prices via Tencent Cloud if other plans do not materialize.

CoreWeave announced at its Fully Connected conference in San Francisco that the Nvidia Vera Rubin NVL72 system is now available on its cloud platform, with Cognition as the first customer, having begun using the systems in early September. Cognition reported up to a 4.8x increase in total token throughput for its SWE-2 inference workloads on the new hardware. CoreWeave also announced plans to offer the Vera CPU as a standalone bare-metal product, with some customers expected to begin testing it in coming weeks.

CoreWeave unveiled CoreWeave Forge, an integrated software platform for training, inference, evaluation, observability, and agent development, at its Fully Connected event in San Francisco on September 30. Available in Free, Pro, and Enterprise editions, Forge includes new tools such as CoreWeave Notebooks, CoreWeave Agent Lens, and RL Rollouts, alongside the generally available CoreWeave Aria AI research assistant. On the hardware side, the Nvidia Vera Rubin NVL72 is now available on CoreWeave's cloud platform, with AI startup Cognition the first customer to run workloads on it, going live in two days. CoreWeave also launched a partner network with early integrations from VAST Data, CrowdStrike, and ClickHouse.

Meta has signed an agreement to lease AI compute capacity from Australian neocloud Firmus across its upcoming Southeast Asia data centers, building on an existing deal in which Meta leases Nvidia GB300 NVL72 compute from Firmus' Melbourne facility. Firmus is co-developing AI factories in the region with operator DayOne, including a Batam, Indonesia site expected to house 170,000 GPUs, and additional facilities in Malaysia where OpenAI is also a customer. The Southeast Asia capacity delivered to Meta will run on Nvidia's full-stack DSX platform, with Firmus' HyperCube modular data center solution integrated into the facilities. Firmus co-founder and co-CEO Tim Rosenfield said the build-out has the scale to support Meta's AI research and model development workloads, and the company's total contracted capacity now exceeds 900MW.

AI data center builder Crusoe has ended a $1.25 billion agreement with Boom Supersonic that would have supplied 29 of Boom's 42-megawatt Superpower natural gas turbines starting in 2027. Boom CEO Blake Scholl confirmed the dissolution on X, noting that Boom still expects to deliver around 250 megawatts of Superpowers to other sites next year and is targeting 1 gigawatt of capacity by 2028. Crusoe spokesperson Andrew Schmitt said the company's energy plans are unchanged and that it will continue using turbines from other suppliers, selected site by site alongside wind, solar, batteries, and grid power. Crusoe's 900-megawatt Abilene, Texas site for Microsoft will use on-site gas turbines, while its 1.2-gigawatt Abilene campus for Oracle and OpenAI runs on grid power with gas turbines as backup.

AMD announced on Monday, September 28, that it will acquire San Francisco-based AI research lab World Labs in an all-stock transaction valued at approximately $8.2 billion, with the deal expected to close by the end of 2026 pending regulatory approvals. World Labs develops spatial intelligence world models that use text, images, and video to generate and simulate interactive 3D environments; its co-founder and CEO, AI researcher Fei-Fei Li, will join AMD as executive vice president and chief scientist. AMD CEO Lisa Su said the acquisition is aimed at co-designing hardware, software, and systems for emerging AI applications including robotics and physical AI. AMD estimates the physical AI market could reach $200 billion by 2035, according to Futurum Group research director Brendan Burke.

Nscale, an Nvidia-backed AI infrastructure provider, filed for a U.S. IPO on September 18 and disclosed four GPU services agreements with Anthropic signed August 25 that could generate approximately $44.6 billion in aggregate payments. The agreements call for Nscale to deliver dedicated infrastructure built around Nvidia Vera Rubin NVL72 systems at the planned Monarch Compute Campus in Mason County, West Virginia, a site spanning roughly 2,250 acres with more than 8 GW of gross planned power capacity. Nscale disclosed it had not yet secured binding financing commitments for the GPU equipment and data center infrastructure required under the agreements. Separately, Anthropic also signed a roughly $35 billion cloud computing agreement with Lambda for infrastructure at a 350-MW facility being developed by Hut 8 in Nueces County, Texas.

AI company Anthropic has agreed to lease compute capacity from Akamai in a seven-year deal valued at $11.6 billion, with a potential expansion clause that could bring the total commitment to approximately $20 billion. Anthropic will use Akamai Cloud's distributed infrastructure and software to support its CPU workloads. Total capital expenditures related to the deal are estimated at approximately $5.5 billion, including roughly $1.7 billion in additional 2026 spending to pre-purchase supply chain components such as memory. Akamai also issued Anthropic a warrant for up to approximately five percent of its common stock, with vesting tied to the expansion of the agreement.

At least 75 U.S. data center projects worth roughly $130 billion were blocked or delayed in the first quarter of 2026, nearly equal to the total for the prior full year, as active opposition groups climbed from 396 to 833 across 49 states in that same period. Underneath local objections over water and power lies a deeper question about who owns compute, since revenue flows to companies headquartered in Seattle, Redmond, or Menlo Park while counties absorb costs including substation strain, truck traffic, and tax incentives forfeited in advance. Technical developments are beginning to loosen the requirement for massive co-located clusters: a 2023 Google DeepMind finding showed geographically separated machines could train models while exchanging roughly 500 times less data, and a March project called Templar trained a 72-billion-parameter model across more than 70 contributors over ordinary internet connections. Nvidia also introduced Spectrum-XGS Ethernet in August 2025 specifically to stitch sites together across long distances as individual facilities hit power and capacity ceilings.

Abandoned cloud workloads, often called zombies, have long consumed wasted resources in data centers, but the GPU era has dramatically raised the financial stakes. IDCA chief research officer Roger Strukhoff said IDCA research indicates that as much as 13% of US cloud usage comes from zombie workloads, while FinOps vendors including Broadcom and AWS estimate overall cloud waste at 25% to 30% or more. Graziano Castro, a developer relations engineer at AI optimization platform maker Akamas, noted that GPU inefficiencies that were once rounding errors on a cloud bill have become very large costs, saying the cost of ignoring inefficiency rose by an order of magnitude almost overnight with the LLM era. Tools from vendors including Google, Flexera, Datadog, and IBM are being used to detect and decommission idle assets, though AI-specific workloads on Kubernetes present new challenges still in early stages of being addressed.

Electricity consumption from data centers has grown 12 percent per year over the last five years, and AI training workloads are expected to drive a substantial further increase, pushing developers toward gigawatt-scale facilities with compressed delivery timelines. The analysis, published by Data Center Dynamics, argues that meeting this demand requires integrating power, cooling, transmission, water, and digital systems as a single industrial campus rather than treating them as separate procurement decisions. Where grid interconnection timelines delay projects by years, behind-the-meter generation such as gas turbines can accelerate deployment, though at higher capital and operational costs. Thermal energy storage is identified as one mechanism to shift cooling demand away from peak pricing periods, reduce demand charges, and support utility demand-response programs without interrupting operations.

Shares of companies tied to AI data center buildout fell sharply on Monday after Anthropic CEO Dario Amodei proposed slowing the pace of frontier model development over the weekend. GE Vernova dropped nearly 9%, Vertiv fell close to 8%, Caterpillar sank more than 4%, and Oracle slipped almost 4%. RBC Capital Markets equity analyst Rishi Jaluria told CNBC that a slowdown in model development and training would likely weigh on Oracle's cloud infrastructure business. Separately, two unnamed sources said hyperscalers are rushing to secure AI debt over the next six weeks, with Amazon raising roughly 4.25 billion pounds last week and Alphabet raising about $10 billion in a euro bond sale in May, though new deals this fall are expected to price at significantly higher rates.

SemiAnalysis has published what it describes as the first verified agentic inference results for NVIDIA's Vera Rubin NVL72 platform, measured using its AgentX benchmark across a fleet of thousands of chips. At 170 tokens per second, Vera Rubin NVL72 delivered approximately 67x the total throughput per TCO compared to the GB300 Dynamo configuration under owning-cost assumptions, and achieved up to 7x better token throughput per megawatt on pre-release software. The benchmark has been validated by major compute buyers including Google Cloud, Microsoft Azure, Oracle, and Meta, and supported by frameworks including vLLM, SGLang, and PyTorch. SemiAnalysis also estimates that even on early software builds, Rubin can earn over 2x more profit per gigawatt than the Blackwell platform, with that gap expected to widen as the software stack matures.

The National Institutes of Health's All of Us Research Program illustrates how much computing power modern medical research already demands. In its June 2026 data release, the program made genomic and health data from more than 747,000 participants available to registered researchers, bringing total enrollment past 883,000. The dataset now includes more than 535,000 whole genome sequences, 1.3 billion genetic variants, and nearly 482,000 linked electronic health records, all housed on a secure Researcher Workbench built on Google Cloud Platform. “There’s a paradox at the heart of precision medicine,” NIH Director Jay Bhattacharya said in announcing the release. “To tailor treatments to individuals, you actually need very large populations to uncover the patterns that connect genetics, lifestyle, and the environment to health outcomes.”

That kind of population-scale genomic analysis runs on infrastructure most people never see. Biowulf, the NIH’s in-house Linux computing cluster, is described by the agency’s Center for Information Technology as the world’s most powerful supercomputer dedicated solely to biomedical research. It has grown to more than 100,000 processor cores and 60 petabytes of storage, is used by roughly 75% of NIH principal investigators, and in fiscal year 2024 alone consumed more than 1 billion core-hours and 6 million GPU-hours. The system helped the Telomere-to-Telomere Consortium publish the first complete, gapless human genome sequence in 2022, and NIH says COVID-19 research on Biowulf consumed more than 87 million CPU hours and produced more than 50 peer-reviewed publications.

That demand is emerging just as the broader market for AI compute tightens. Apollo Global Management’s wealth insights team described on-demand GPU capacity as “effectively sold out” in a June 2026 analysis, noting that rental rates for Nvidia H100 GPUs rose from roughly $1.85 an hour in late 2025 to about $2.40 an hour by March 2026, while spot prices for high-bandwidth memory chips climbed roughly eightfold since early 2025. Apollo named healthcare specifically as one of the sectors “only beginning to deploy AI at scale” just as that squeeze sets in, alongside legal services and financial analysis. Data Center Knowledge reported the binding constraint has shifted from the power shortages that dominated 2024 and 2025, when Microsoft CEO Satya Nadella said the company had “a bunch of chips sitting in inventory that I can’t plug in,” to semiconductor manufacturing capacity itself. “Silicon is the binding short-term constraint. Power is the binding long-term constraint,” HyperFrame Research’s Stephen Sopko told the outlet.

Drug makers with the capital to do so are responding by building and owning compute capacity rather than competing for it on the open market. Recursion Pharmaceuticals brought online BioHive-2, a supercomputer built with Nvidia using 504 H100 GPUs delivering 2 exaflops of AI performance, which debuted at No. 35 on the TOP500 list of the world’s most powerful supercomputers. Recursion’s chief technology officer said the company can now get “80% of the value with 40% of the wet lab work” that traditional drug discovery required. Roche has gone further, announcing an AI factory spanning more than 3,500 GPUs across facilities in Europe and the United States, built on Nvidia’s newest Blackwell chips and expected to be fully operational by early 2027. “Everybody wants to get their hands on Nvidia chips,” biopharma AI consultant Christian Hein told SWI swissinfo.ch.

Hospitals, by contrast, mostly cannot build their own supercomputers and instead depend on outside vendors and cloud infrastructure for the same AI capacity pharma companies are racing to own. In 2026, the U.S. Department of Veterans Affairs began scaling its ambient AI scribe tool from a 10-site pilot to all of its more than 130 medical centers nationwide. Mount Sinai Health System integrated OpenEvidence’s AI clinical decision support directly into its Epic electronic health record system across all seven of its hospitals. HonorHealth rolled out Abridge’s ambient documentation platform to roughly 3,000 physicians and advanced practice providers, and the University of Texas System’s earlier pilot with Qualified Health generated more than $15 million in run-rate financial impact within six months before expanding across all eight of its health institutions. Each of these systems now depends on rented, high-capacity cloud infrastructure to keep those tools running, in a market where that same infrastructure is getting harder and more expensive to secure.

That dependence is starting to show up in how healthcare organizations plan their own facilities, too. Commercial real estate firm Colliers has noted that hospitals and health systems increasingly compete with data center developers for the same scarce inputs, reliable electricity, available land, and network connectivity, while their own IT infrastructure has to support higher computing density and GPU-ready systems that older hospital data centers were never built for. For an industry historically organized around clinical staffing and bed capacity, the amount of processing power behind a diagnosis, a drug candidate, or a documentation tool is quickly becoming a resource healthcare organizations can no longer take for granted.

Schneider Electric has introduced 2.5-megawatt power distribution modules specifically designed for AI data center deployments, where rack densities and power loads are outpacing equipment designed for conventional enterprise infrastructure. The modules are intended to simplify power delivery at scale, reducing installation complexity for operators building out GPU clusters. Schneider did not disclose pricing or initial customer commitments in available reporting.

A coalition warned that AI data center buildouts are creating acute pressure on global memory supply, according to Data Center Knowledge. The group flagged that demand for high-bandwidth memory and DRAM used in AI accelerator systems is outpacing production capacity, potentially creating bottlenecks that slow deployment timelines. The warning adds a supply-chain dimension to existing concerns about GPU and power availability.

Microsoft's AI data center campus in Wisconsin is now fully operational, marking the completion of a major hyperscaler buildout in the Midwest. The facility represents one of the largest AI infrastructure deployments in the region. Full activation means the campus is now processing live workloads at scale, shifting from construction to revenue-generating operations.

South Korean construction companies are reporting that their experience building AI data centers has forced them to develop in-house AI and engineering capabilities they say are now core to their business model, according to 아시아경제. Firms described a shift away from traditional construction methods toward more technology-intensive processes driven by the precision demands of hyperscale data center projects. Executives said the transformation is irreversible and is reshaping workforce and project management practices.

The Modern War Institute at West Point published an analysis framing AI data centers as critical military terrain that adversaries would prioritize in future conflicts. The piece examines the concentration of AI training and inference infrastructure in a small number of large campuses as a strategic vulnerability, arguing that military planners must account for their defense. Authors call for hardening, redundancy, and distributed architecture considerations to be integrated into national AI infrastructure planning.

A new report projects Microsoft is targeting 38 gigawatts of total data center capacity by 2032, a figure that would represent a massive expansion of the company's global infrastructure footprint. The scale implies sustained capital expenditure across multiple continents over the next six years. No single facility or region was identified as the primary driver of that total.

Oracle delivered 300,000 GPUs during its first fiscal quarter of 2027, according to Data Center Dynamics. The delivery volume reflects the company's accelerating build-out of cloud AI infrastructure as it competes with larger hyperscalers for enterprise AI workloads. Oracle had previously issued an RFP for 2GW of new renewable capacity in New Mexico, signaling continued expansion plans.

SemiAnalysis published an analysis of Nvidia's backstop agreements with cloud providers and AI labs, examining how the GPU maker structures supply deals so that demand risk is distributed across customers rather than absorbed by Nvidia itself. The piece details how these arrangements insulate Nvidia from cancellation losses while leaving counterparties exposed if AI workload demand softens. No specific dollar figures were disclosed in the snippet, but the analysis covers agreements spanning multiple hyperscalers.

Finnish company 3 E Network has released a blueprint for a new AI data center in Mikkeli, Finland, designed specifically around NVIDIA's Vera Rubin GPU architecture. The facility is engineered to support the high power densities and interconnect demands that Vera Rubin systems require, positioning it among the first campuses publicly tailored to that next-generation chip platform. No construction timeline or capacity figure was disclosed in the announcement.

The joint venture between Google and Blackstone built around Google's TPU chips has been formally named Crux AI. The company has hired Alan Duong, previously head of data center engineering at Meta, to lead its infrastructure operations. Crux AI is positioning itself as a neocloud provider offering TPU-based compute capacity to enterprise and AI customers.

NVIDIA is expanding its DSX platform into broader data center infrastructure functions, moving beyond GPU compute into networking, storage, and systems management layers. The push represents an effort by NVIDIA to capture more of the infrastructure stack as hyperscalers and colocations build out AI capacity. Specific product and partnership details were outlined at a recent industry event.

Google announced a two-year €13 billion investment in AI infrastructure in Finland, the company confirmed in a Google Cloud press release. The commitment covers data centers and related digital infrastructure and represents a significant expansion of Google's European AI compute capacity. Finnish officials have welcomed the investment as a major economic development for the country.

SemiAnalysis has published a technical analysis examining where robotic AI systems should perform inference, comparing on-device processing against datacenter-based computation for latency, cost, and reliability trade-offs. The analysis finds that the answer depends heavily on task type, connectivity, and acceptable latency thresholds, with neither model dominant across all use cases. As robotics becomes a major new category of AI deployment, the infrastructure implications for both edge hardware and centralized data centers are significant.

Data Center Frontier published an analysis arguing that AI data centers are no longer passive power consumers but are being designed and operated as active components of broader energy systems. The piece examines how hyperscalers and large colocation operators are integrating on-site generation, storage, and demand-response capabilities to manage grid interconnection constraints. The shift reflects pressure from utilities and regulators who want large loads to contribute to grid stability rather than simply draw from it.

Anthropic committed to $517 billion in compute agreements over the past 11 months, according to Data Center Dynamics. The scale of these contracts reflects the AI lab's aggressive infrastructure buildout as it competes with OpenAI and Google. The figure covers cloud and hardware procurement deals across multiple providers.

SemiAnalysis reports that Google is moving forward at full speed with externalizing its TPU inference capacity through a platform called InferenceX, making its custom chips available to outside customers. The move marks a strategic shift from using TPUs exclusively for internal workloads to competing directly with Nvidia in the inference-as-a-service market. Details on pricing and initial customers were included in the SemiAnalysis analysis.

PwC has published an analysis projecting that $31.6 trillion in capital expenditure will flow through AI-era infrastructure investment globally. The report maps where that spending concentrates, covering data centers, semiconductors, energy systems, and network infrastructure. PwC frames this as a structural economic shift rather than a cyclical investment cycle.

The National Institute of Standards and Technology has released guidance on securing AI data center architecture, covering security posture assessment and emerging standards applicable to facilities running large-scale AI workloads. The publication addresses how the physical and logical structure of AI data centers differs from conventional facilities and what those differences mean for security controls. NIST guidance typically informs federal procurement requirements and is widely adopted as a baseline by private sector operators.

Equinix, Together AI, and Nvidia have announced a joint partnership to launch a product called Inference Exchange, which is designed to give enterprise customers streamlined access to AI inference capacity across Equinix's global data center footprint. The arrangement combines Equinix's colocation infrastructure, Together AI's model serving software, and Nvidia's GPU hardware. Financial terms were not disclosed.

ChatGPT, Claude, and Grok experienced simultaneous service outages, according to Data Center Dynamics, in an event that briefly disrupted access to three of the most widely used AI platforms. The coincidence of outages across separate operators raises questions about shared infrastructure dependencies or common upstream failures. No single cause has been publicly confirmed by all three companies.

Data Center Knowledge reports that self-improving AI systems, which retrain or refine themselves autonomously, could drive compute demand well beyond current forecasts and strain data center capacity in ways that static workload models do not capture. The analysis highlights that recursive improvement cycles generate unpredictable bursts of training demand rather than the steady growth operators plan for. Capacity planners are warned that traditional utilization curves may become unreliable.

Data Center Knowledge identifies fiber network capacity as a potential constraint on data center expansion, arguing that the industry's focus on power and land overlooks limits in high-density fiber routing to and within large campuses. The analysis points to long lead times for new fiber builds and permitting hurdles for buried conduit as compounding factors. Operators racing to deliver AI capacity may find connectivity, not just power, delaying commissioning timelines.

Amazon is planning a new subsea cable system linking the United States and Japan, Data Center Dynamics reports. The cable would expand Amazon Web Services' transpacific network capacity, supporting data center operations and cloud traffic on both ends of the route. Subsea cable investments of this scale typically involve hundreds of millions of dollars in capital expenditure. The project would add redundancy and throughput to one of the world's highest-traffic data corridors as AI workloads increase demand for low-latency international connectivity.

AI inference demand is reshaping data centre infrastructure requirements across networking, storage, power delivery, and cooling, according to ET Datacenters, as inference workloads differ substantially from training in their latency sensitivity, traffic patterns, and hardware configurations. Operators are finding that infrastructure optimized for batch training runs cannot efficiently handle real-time inference at scale, driving significant retrofitting and new-build specification changes. The shift is accelerating procurement cycles for specialized accelerators and high-bandwidth networking equipment.

Data center and power company Iren has secured a new contract with what it described as a leading frontier AI lab, confirming it is on track to fully exit the cryptocurrency mining business by the end of 2026. The company did not name the client, but the deal marks a strategic pivot from crypto hosting toward AI compute workloads. Iren's transition reflects broader industry pressure on crypto-adjacent operators to reposition their infrastructure for AI revenue.

IREN, formerly Iris Energy, has made its first delivery of AI cloud computing capacity to Microsoft under a contract valued at up to $9.7 billion. The milestone represents the initial phase of one of the largest AI infrastructure supply agreements between a specialist compute provider and a hyperscaler. Further capacity tranches are expected to follow as IREN expands its GPU data center footprint.