News on the plumbing that powers modern AI: compute and GPUs, inference and training platforms, cloud partnerships, and the supply chain behind the models practitioners use every day.
Stories
1,530
Latest source update
August 2, 2026
Coverage
Live
Topic brief
What to know about AI Infrastructure
Brief updated Aug 2, 2026
AI infrastructure covers the physical, software and financial layer beneath AI models and agents: data centers, power procurement, chips and packaging, high-bandwidth memory, storage, interconnects, the cloud and networking systems that connect them, and the compilers, runtimes, routers and gateways that turn that hardware into served tokens. It is built and financed by hyperscalers such as Meta, Google, Microsoft and Amazon, chipmakers such as NVIDIA, AMD, Intel and Qualcomm, foundries and suppliers such as TSMC, Samsung and SK hynix, and capital providers such as BlackRock, Brookfield and large infrastructure funds.
For practitioners, training and serving frontier models depends on gigawatt-scale power commitments, multi-year chip and packaging supply agreements, and memory capacity negotiated years in advance. Delays or bottlenecks in any of these layers, whether in rack manufacturing, advanced packaging, utility interconnection or export controls, directly constrain what compute is actually available to ML teams regardless of model capability. Memory is the clearest current example: data center demand pulls supply away from other markets and sets prices well before a model is trained.
The financing structure has become part of the technical stack. Capacity is increasingly built on vendor guarantees, equity commitments from chip suppliers to their own customers, and long-term leases, which means counterparty risk, lease terms and delivery milestones now shape schedules as much as fab capacity does. That layer also intersects with macroeconomics, national industrial policy and local politics: AI capital spending is large enough that hyperscaler capex guidance draws close investor scrutiny, sovereign programs are negotiated at national level, public money and state incentives are being committed to private campuses, and hyperscale projects draw environmental review, utility tariff design and community-benefit negotiations at the state and municipal level. Practitioners planning large deployments need to track energy, grid and regulatory conditions alongside chip roadmaps.
What changed recently
The newest headline is a number that has to be unpacked before it can be used. Yonhap, citing presidential policy chief Kim Yong-beom, reported that participants at a July 24 summit in San Francisco hosted by President Lee Jae Myung agreed to pursue semiconductor and AI infrastructure cooperation projects worth a combined $950 billion. That figure combines separate supply, development and infrastructure plans; it is not a single purchase. Its largest disclosed piece is a long-term arrangement under which SK Group would supply $750 billion in high-performance semiconductors to global technology companies including Nvidia, and Yonhap also reported planned joint investment in large AI data centers with a combined capacity of 5 gigawatts and 2 million GPUs, without establishing how much of that infrastructure sits inside the $750 billion or when any component would become operational. The other major disclosed component is Samsung's own announcement of a memorandum of understanding with Broadcom across memory, foundry and advanced packaging, which Samsung said the companies estimate at more than $200 billion over five years through 2030, covering high-bandwidth memory for Broadcom's next-generation AI accelerators and Samsung foundry processes at 2 nanometers and below. Read all of it as agreements to pursue work, not booked revenue. The results underneath are firmer: Samsung reported 89.5 trillion won and SK Hynix 60.5 trillion won of June-quarter operating profit, about 150 trillion won combined, with SK Hynix saying HBM4 mass shipments began during the quarter and that it completed long-term agreements with around 10 customers, and the Associated Press reporting that Samsung had secured long-term supply contracts with five major global data-center customers. On the buying side, Andy Jassy raised Amazon's expected 2026 cash capital spending to roughly $220 billion from the $200 billion planned in February, citing higher memory costs as the main reason, and still told investors Amazon will not have enough capacity to meet all of its 2026 demand and expects the same in 2027.
The second new item shows what it now takes to keep power on a site. SpaceXAI said on July 30 that an agreed order with the Mississippi Department of Environmental Quality requires all 69 temporary gas turbines at its Southaven facility to be removed by July 2027, with removals able to begin as early as August 2026, and says it is bringing online a 1.2 GW natural-gas plant with 41 permanent turbines built under a Clean Air Act permit; MDEQ's public database lists that air-construction permit as issued to MZX Tech on March 11, 2026. TechCrunch and Data Center Dynamics report that the NAACP and environmental-law groups challenged xAI and MZX Tech over the temporary units, and the retrieved sources do not show that the agreed order resolves that dispute. Permission is a scarce input across this cycle: New York's Executive Order 62 holds incomplete state permit applications for data centers capable of consuming at least 50 megawatts while the Department of Public Service prepares a Generic Environmental Impact Statement, a pause the governor's office says can last up to one year, and Pennsylvania's data-center fight has moved into 2026 campaign advertising. Capital, by contrast, keeps finding new structures: Meta disclosed $278.99 billion of not-yet-commenced leases as of June 30, about 53% above the $182.88 billion reported at March 31, plus roughly $68 billion of additional data-center leases signed in July; BlackRock-managed funds will own 80% of the roughly $14 billion, 1 GW El Paso campus that Meta will lease back; and CNBC reported that Nexus Data Centers is seeking roughly $15 billion for an Anthropic-linked campus in Hubbard, Texas, with Google reported to have agreed to backstop Anthropic with its investment-grade credit rating. At the software layer, Qualcomm said it completed its acquisition of Modular, valued by Axios at $3.9 billion in stock, Nscale signed a definitive agreement to acquire Anyscale on undisclosed terms, and OpenAI cut GPT-5.6 Luna prices by 80% and Terra by 20% on July 30 while leaving Sol unchanged, after saying Sol-assisted kernel work had reduced its own end-to-end serving costs by 20%.
What to watch
Dated and checkable items in this evidence: SpaceXAI says removals of the 69 temporary Southaven turbines can begin as early as August 2026 and must be complete by July 2027, and the retrieved sources do not show that the agreed order resolves the Clean Air Act challenge brought by the NAACP and environmental-law groups; Samsung describes its Broadcom arrangement as an MOU for planned work over five years through 2030, so watch for delivery milestones rather than movement in the headline total. EuroHPC's AI Gigafactories tender closes November 12, 2026, with selection expected in early 2027 and operations targeted within 18 months of selection. New York's hold under Executive Order 62 runs until the Department of Public Service submits its final Generic Environmental Impact Statement, up to one year, and applications DEC deemed complete before July 14 sit outside it. Nscale's acquisition of Anyscale is expected to close in the second half of 2026 subject to conditions and regulatory approvals, with terms undisclosed and the $1.65 billion figure in secondary coverage unconfirmed. Snowflake says Cortex AI Gateway will enter public preview soon, with several identity integrations still planned for private preview. Nexus's roughly $15 billion financing has not been announced as a completed deal, and Octus reported a bridge expected to be replaced by bonds or loans in the fall. Cash interest on Galaxy Helios' $3.507 billion of 9.875% notes is scheduled to begin February 1, 2027, with the first principal amortization due at least 10 months after the project is completed. Meta's roughly $68 billion of July leases are expected to commence in 2027 and 2028, and the El Paso venture expects to begin bringing capacity online in 2028. SK Telecom's up-to-2-gigawatt AI factory targets a 2027 first phase, while NVIDIA's planned $1 billion for the NAVER GAK Sejong expansion to 200 MW by 2028 remains subject to NAVER securing at least $9 billion of committed financing separately. Aalo and Crusoe target a 2027 proof of concept at Idaho National Laboratory and state an intention to deploy 50 MWe Pods by the end of 2029. The DOE Paducah power service agreement still requires Kentucky Public Service Commission approval; Mississauga's interim control bylaw has only been directed for preparation; Australia's proposed data-centre standards still have no set coverage threshold and face Queensland and Northern Territory opposition to a renewable-power requirement; Boulder City Council voted on July 14 to appeal BLM's Townsite approval; and NVIDIA's published Vera Rubin specification still lists 54TB of LPDDR5X per NVL72 rack despite the reported SOCAMM reduction.
Comparison
caveat
figure
company
what it covers
Up from the $200 billion plan announced in February and includes spending beyond AWS such as robotics, semiconductors and satellites, with technology and AI accounting for most of the increase; Jassy cited higher memory costs as the main reason and said Amazon still will not have enough capacity to meet all of its 2026 demand
About $220 billion
Amazon
Expected 2026 cash capital expenditures, stated by CEO Andy Jassy on the earnings call as Amazon reported second-quarter results on July 30
Alphabet's SEC-filed release showed $119.8 billion of second-quarter revenue and $24.8 billion from Google Cloud, up 82%; Reuters reported the Nasdaq Composite fell 2.15% on July 23, but also attributed that session to Brent crude settling above $100 a barrel and higher bond yields, not AI spending alone
$195 billion to $205 billion
Alphabet
2026 capital-expenditure guidance, raised from a prior range of $180 billion to $190 billion, as reported by Axios and Reuters
Down from the roughly $190 billion given in April only because a longer estimated useful life will classify more future datacenter leases as operating rather than finance leases; Microsoft says its underlying calendar-2026 investment expectations are unchanged and expects September-quarter capex to exceed $50 billion
About $175 billion
Microsoft
Expected reported calendar-2026 capital expenditures, stated on the July 29 fiscal fourth-quarter earnings call
A quarterly figure rather than annual guidance; free cash flow was $784 million, and the quarter's 14% profit decline also reflected $2.40 billion of charges related to legal proceedings and $1.18 billion of severance, so it cannot be attributed to AI spending alone
$31.08 billion
Meta
Second-quarter 2026 capital expenditures, reported July 29
Frequently asked questions
What does the reported $950 billion South Korea figure actually cover?+
It is an aggregate of separate plans announced around a July 24 summit in San Francisco hosted by President Lee Jae Myung, not a single purchase. Yonhap, citing presidential policy chief Kim Yong-beom, reported that participants agreed to pursue cooperation projects worth a combined $950 billion, including a long-term arrangement under which SK Group would supply $750 billion in high-performance semiconductors to global technology companies including Nvidia. Yonhap also reported planned joint investment in large AI data centers with a combined capacity of 5 gigawatts and 2 million GPUs, and the public details do not establish how much of that infrastructure is included in the $750 billion figure or when each component would become operational. The other major disclosed component is Samsung's memorandum of understanding with Broadcom spanning memory, foundry and advanced packaging, which Samsung said the companies estimate at more than $200 billion over five years through 2030. Because the retrieved materials describe an agreement to pursue projects, a long-term supply arrangement and an MOU, the headline should not be read as booked revenue, a completed purchase, or capacity that exists today.
Is AI compute still supply constrained, or has the buildout caught up?+
On this evidence it is still constrained, and the operator saying so most plainly is Amazon. Andy Jassy told investors that even at roughly $220 billion of 2026 cash capital spending, Amazon will not have enough capacity to meet all of its 2026 demand, expects the same dynamic in 2027, and described demand already visible for 2028 as striking; he cited higher memory costs as the main reason for lifting the forecast. Memory is where the shortage prices itself: Samsung and SK Hynix posted about 150 trillion won of combined June-quarter operating profit and both tied the records to AI-server demand and higher prices, with SK Hynix reporting a 76% operating margin and saying HBM4 mass shipments only began during the quarter. Long-term agreements are spreading, with SK Hynix saying it completed deals with around 10 unnamed customers and the Associated Press reporting that Samsung had secured long-term supply contracts with five major global data-center customers. These are management statements and company-reported results, not independent capacity measurements.
Did Microsoft cut its AI infrastructure spending?+
No, according to the company. Microsoft said on July 29 that its underlying calendar-2026 investment expectations remain unchanged, and that the drop in expected reported capex to about $175 billion, from the roughly $190 billion given in April, comes from a longer estimated useful life for datacenters and office buildings that will cause more future datacenter leases to be classified as operating leases rather than finance leases. Finance leases are included in capital expenditures and operating leases are not, so the roughly $15 billion difference is a classification effect rather than an equivalent reduction in planned infrastructure investment. Microsoft recorded $41 billion of capital expenditures in the quarter ended June 30, including $5.6 billion of finance leases and $35.8 billion of cash additions to property and equipment. It expects September-quarter capex to exceed $50 billion, including the effect of the reclassification, and expects fiscal-2027 capex to grow year over year.
Why is so much data-center capacity moving off the operator's balance sheet?+
The disclosures in this evidence show several vehicles doing the same job. Meta reported $278.99 billion of not-yet-commenced leases as of June 30 and signed roughly $68 billion more in July with 18-to-20-year terms, while placing 80% of its roughly $14 billion El Paso campus with BlackRock-managed funds and leasing the entire site back under a four-year initial term with four extension options and declining residual-value guarantees. Galaxy Digital's Helios unit priced $3.507 billion of 9.875% senior secured notes due 2031 against a CoreWeave-leased Texas project. CNBC reported that Nexus Data Centers is seeking roughly $15 billion for an Anthropic-linked campus in Hubbard, Texas, with Google reported to have agreed to backstop Anthropic with its investment-grade credit rating, while Hyperscale Data said it monetized approximately 100 Bitcoin and established a Bitcoin-backed credit facility for a Michigan campus. In each case, delivery timing, lease terms and counterparty quality become the financing risk, which is why VanEck's estimate that only about 25% of leased AI and HPC capacity had been delivered at the time of its report is a useful reference point.
What is the biggest non-technical risk to a new campus right now?+
Permission and local resources, on this evidence more than chips. New York is holding incomplete state permit applications for facilities capable of consuming at least 50 megawatts for up to a year, though applications DEC deemed complete before July 14 and permits issued by local governments are outside the order. Mississauga council directed staff to prepare an interim control bylaw that would temporarily prohibit new AI, cloud-computing and hyperscale data centres, Saskatoon directed administration to review the tradeoffs before the city pursues proposals, and Australia has proposed national standards requiring large new data centres to underwrite their own new power supply and pay their full share of grid-connection costs. Power permits carry their own tail: SpaceXAI's agreed order with Mississippi regulators sets a July 2027 deadline to remove all 69 temporary Southaven turbines, and TechCrunch and Data Center Dynamics report that the NAACP and environmental-law groups challenged xAI and MZX Tech over those units. Litigation follows the same path elsewhere, with the Greater Birmingham Humane Society suing over a Nebius project on zoning grounds and Imperial Valley Computer Manufacturing suing the irrigation district for water service that KPBS puts at about 260 million gallons of Colorado River water per year. Pennsylvania shows the political version: AP reported a roughly $200,000, two-week television advertisement attacking data center secrecy, tax breaks and electricity costs in a competitive House race.
Is inference actually getting cheaper?+
List prices moved down at the low end of one family. OpenAI cut GPT-5.6 Luna by 80%, to $0.20 per million input tokens and $1.20 per million output tokens, and Terra by 20%, to $2 and $12, on July 30, while leaving Sol unchanged. It attributed the reductions to improvements across model serving and production software, saying Sol-assisted kernel work reduced end-to-end serving costs by 20% and that related experiments increased token-generation efficiency by more than 15%; both are company-reported and not independently validated in the retrieved reporting. OpenAI also said subscription prices and quota budgets for Codex and ChatGPT Work are unchanged but that use of the two cheaper models now consumes fewer credits, and it introduced Fast mode for the API in place of Priority Processing. On the serving side, Amazon Bedrock lists a 272K-token context window across the GPT-5.6 tiers with prompt caching carrying a reported 90% cached-input discount, and ScaleFlux introduced an SSD platform for KV-cache offload beyond GPU HBM and host DRAM, designed for a vendor-reported 7 to more than 10 effective drive writes per day at five years for those workloads. Note that the discounts of as much as 97.8% below official list pricing documented by Vectoral come from third-party relay services that pool and resell access to U.S. models, not from official rates.
+1
