How Much Are AI Companies Spending on Data Centers?

By Muninder Adavelli. Scraping, SEO, and proxies for over a decade.

The 2026 AI-infrastructure numbers are the largest coordinated capital deployment in the history of commercial computing, and they can't be added up because they aren't measuring the same thing. Goldman Sachs Research puts global AI investment at about $1 trillion in 2026 alone. McKinsey's "Cost of Compute" analysis projects $5.2 trillion of cumulative AI-related data-center capex through 2030. A Reuters investigation found $1.09 trillion of "uncommenced" lease commitments sitting off Big Tech balance sheets. All three are correct. All three describe different overlapping slices of the same buildout. This is a look at each of the four buckets (hyperscaler capex, AI-native compute commitments, chip purchases, and lease pipelines), and how the seven companies driving the spend are dividing $1–7 trillion of infrastructure between them. A pool of 868 stats and numbers was cross-checked and cross-tabulated by hand for this piece; sources are cited at the foot.

1. The 2026 numbers, and what each of them is actually counting

1.1 Goldman Sachs: $1 trillion of AI investment globally in 2026

Goldman Sachs Research estimates AI investment will total roughly $1 trillion globally in 2026, with just under $600 billion of that inside the United States. Cumulative investment through the end of the year: about $1.8 trillion. Goldman's economists arrive at that number by augmenting the standard hyperscaler-capex measurement in four specific ways: adding capex projections for public non-hyperscaler US companies in Goldman's AI-related equity baskets, capturing media-tracked capex for pivotal private companies, extending the model to AI-exposed non-US companies, and subtracting off 2022 baseline capex on the view that nearly all incremental capex since 2022 has been allocated toward AI projects. Goldman also excludes financial leases where possible to avoid double-counting hardware capex that shows up on multiple companies' books. What matters for a reader: Goldman's $1T is a broader definition than "hyperscaler capex alone." It's investment across the AI supply chain, deliberately drawn wide.

1.2 McKinsey: $5.2 trillion of AI-related capex through 2030

McKinsey's "Cost of Compute" analysis is the source of the widely-quoted "$7 trillion race" framing. The base-case forecast breaks down as $5.2 trillion of cumulative AI-related data-center capex through 2030, plus another $1.5 trillion for non-AI workloads; $6.7 trillion combined. McKinsey publishes two alternative scenarios: a constrained-demand path at $3.7 trillion (78 GW of new AI-related capacity added between 2025 and 2030), and an accelerated path at $7.9 trillion (205 GW added). Where the $5.2T lands is more interesting than the headline. 60% of it ($3.1 trillion) goes to technology developers and designers, meaning chips and computing hardware. 25% ($1.3 trillion) goes to "energizers," meaning power generation, transmission, cooling, and electrical equipment. The remaining 15% ($0.8 trillion) goes to builders (land, materials, and site development). That split is important; it's the underlying grammar of the $1T-plus 2026 print.

1.3 PwC: cumulative AI capex is a multi-decade curve, not a one-year print

PwC's 2026 Global AI Infrastructure Outlook frames the buildout as a decades-long compounding trajectory rather than a single-year spike. Central forecast: $31.6 trillion cumulative global investment in AI infrastructure through 2050, with Asia Pacific alone accounting for $8.2 trillion, led by China and India. A trade-policy-and-export-control shock scenario cuts the total to $25.5 trillion, a $6 trillion divergence that hinges on geopolitical stability. PwC identifies five factors that will determine where investment lands: power availability (the hardest constraint), connectivity, security, policy certainty, and community consent, plus GPU access as a cross-cutting geopolitical variable. Unlike single-year prints, the PwC frame emphasises that on an annualised basis data-centre capex rises from roughly $800 billion per year in 2026 to $1.8 trillion per year in 2050. The 2026 numbers are the early phase of a decades-long compounding curve, not a peak.

1.4 The Reuters lease-burden discovery: $1.09 trillion already committed, off balance sheet

Reuters' August 2026 investigation found that the five biggest US technology companies have collectively committed roughly $1.09 trillion in "uncommenced" lease agreements: data-center facilities under construction or contracted for future occupancy, which haven't yet appeared on balance sheets as recognised lease liabilities. Oracle alone disclosed $260 billion of uncommenced commitments, the largest concentration among the group. Meta disclosed $278.99 billion, Amazon $137.21 billion, and Alphabet $85.2 billion. The pipeline for the five companies climbed to about $1.16 trillion after later agreements. Uncommenced lease commitments are typically undiscounted, spread over many years, while recognised lease liabilities reflect present value, so the $1.09T isn't directly comparable to debt figures. It's arguably the truest measure of committed 2026-and-beyond AI infrastructure spending, and it's the number nobody had been aggregating as either capex or debt.

2. Google: $205 billion of 2026 capex, up from $85 billion of guidance a year ago

2.1 The $195–205B guidance and how it got there

Alphabet raised its 2026 capital-expenditure guidance to a $195–205 billion range, from a prior $85 billion figure just twelve months earlier, and above an intermediate $180–190 billion guidance the company had been operating with. CFO Anat Ashkenazi cited AI-driven demand as the accelerant. Context on the base: Alphabet's 2024 capex was $91.45 billion, up from $52.5 billion in 2023. So the $205 billion 2026 number represents roughly 2.2× the 2024 print, and Alphabet's ability to fund it from operating cash flow (without new debt issuance) is one of the reasons Google's spending profile looks structurally different from OpenAI's or CoreWeave's, both of which rely on external financing.

2.2 The Finland deal: $15.1 billion, bring-your-own-power, and a 22-year nuclear PPA

Google announced in September 2026 that it will invest at least €13 billion ($15.1 billion) in AI infrastructure in Finland over 2027 and 2028, described as Google's "single biggest investment in Europe." The spending will cover data centres, electricity-grid improvements, and clean-energy and battery projects, and is projected to contribute roughly €3.6 billion to Finland's GDP during the construction phase and support 7,000 jobs annually once operational. The novel structural piece is the power arrangement. Alongside the buildout, Google signed a 22-year purchase agreement for up to 50% of the energy output of one of Finland's two nuclear plants, operated by Fortum. Google's President and Chief Investment Officer Ruth Porat described the model as "BYOP" (bring your own power). Fortum and Google will also explore new nuclear and renewable development. Finland is not an isolated case; per Reuters' broader coverage, US hyperscalers have been expanding investments in markets like Finland, Germany, and Britain, where a mix of cool climate, low-carbon energy, and grid capacity meets the AI buildout's specific requirements.

2.3 The Intersect acquisition and Google as banker to Anthropic

In December 2025 Alphabet completed the $4.75 billion acquisition of data-centre-infrastructure firm Intersect. Alphabet has said little publicly about the acquisition's operational integration, but the deal itself signals the direction of the 2026 capex playbook: hyperscalers are increasingly moving upstream, buying not just facilities but the companies that supply specialised infrastructure. The acquisition preceded the 2026 capex raise, suggesting Google's infrastructure strategy was being restructured before the guidance jump was announced. Separately, Financial Post and FT reporting details Google's emergent role as banker to Anthropic through preferred-stock investments, cloud commitments, and structured financing arrangements. The template is being called the "Wall Street finance machine" for how a hyperscaler underwrites an AI-native customer that also spends billions with the hyperscaler. It's a two-sided model: Google gets Anthropic's cloud revenue and increases its own utilisation; Anthropic gets scaled compute plus growth capital. The finance-machine framing is likely to reappear elsewhere in the sector.

3. Microsoft: $80 billion committed to AI data centers through fiscal 2026

3.1 The $80B figure and what it covers

Microsoft has committed $80 billion to AI data-centre investment through fiscal year 2026, per DataCenters.com's report describing the buildout as the "digital backbone for a multimodal world." The $80B covers both new construction and expansion of existing Azure regions; roughly half is targeted at US facilities. Framed relative to the broader analyst consensus, JPMorgan (Morgan Stanley uses this same figure in market notes) estimates 2026 capex for the five largest US hyperscalers at about $697 billion, and $800 billion is the number that recurs across other analyst reports for the broader hyperscaler group. Microsoft's $80B is a share of that total, and the DataCenters.com single-source status means it should be read as Microsoft's own guidance conveyed through that outlet rather than a cross-verified figure.

3.2 Purpose of the buildout: multimodal AI at scale

The DataCenters.com framing positions the $80B as infrastructure for concurrent workloads across text, image, video, and code models, meaning capacity for AI applications that aren't fully in production yet. The demand justification is anchored in Microsoft's OpenAI partnership, which drives a significant portion of the projected compute load. The specific terms of that partnership remain undisclosed in the public pool, and speculation about the split between OpenAI-driven demand and Microsoft's own Copilot, Azure, and enterprise workloads would exceed what the reporting supports. What is clear is scale: $80 billion is Microsoft's largest single-year infrastructure program to date, and it lands in the same 2026 window as OpenAI's Stargate and Anthropic's $50B American AI infrastructure announcement, meaning the compute pipeline for the three flagship AI-native companies is landing simultaneously.

4. Meta: $130–145 billion of 2026 capex, $600 billion committed to the US over three years

4.1 The $130–145B 2026 capex forecast

Meta raised its 2026 capital-expenditure forecast to a $130–145 billion range, a step up from $115–135 billion at the start of the year and $125–145 billion in April: three sequential revisions upward inside a single fiscal year. CEO Mark Zuckerberg framed the buildout as "aggressively front-loading capacity so we're prepared for the most optimistic cases." Meta's free cash flow for the most recent quarter hit its lowest level in five years at $784 million, compared to $8.5 billion in the same quarter a year earlier: a sharp compression driven by the capex step-up. That's the direct signal. Meta is trading current cash for capacity, and doing so at a scale that meaningfully lowers reported free cash flow while the company remains committed to the ramp.

4.2 The $600B US pledge: what Zuckerberg said, what it means

Separately from capex, Meta announced in November 2025 that it will invest at least $600 billion in the US over the next three years, including AI data centres. Zuckerberg first made the pledge publicly at a White House dinner in September 2025 with President Trump. The $600B is a multi-year US-investment envelope covering infrastructure and jobs, of which AI data centres are one component. It is not the same as capex. The distinction matters because "$600 billion" is the number that will circulate publicly, and the actual on-balance-sheet AI-data-centre spending inside it is captured by the $130–145 billion 2026 capex figure and its 2027–2028 successors. Meta's positioning is deliberate: the $600B frames the company as a national infrastructure partner, while the $130–145B is the accountable line item.

4.3 The $27B Louisiana data centre with Blue Owl Capital

In October 2025 Meta closed a $27 billion financing deal with Blue Owl Capital to fund its Louisiana data centre, Meta's biggest project globally. The structure is significant: rather than paying from operating cash, Meta partnered with a private-credit provider to fund a single site, moving the financing off Meta's balance sheet into the partner structure. It's a template for how hyperscalers are using structured private credit for individual mega-sites, and it foreshadows the industry's broader shift toward chip- and infrastructure-backed lending discussed in section 9 below.

5. OpenAI: $300 billion signed with Oracle, $280 billion by 2030, $500 billion in Stargate, and a $750 billion running total

5.1 The Oracle contract: $300 billion over about five years

OpenAI's largest single committed spend is its Oracle contract, reported as one of the biggest cloud deals ever signed, with OpenAI expected to buy roughly $300 billion in computing power from Oracle over about five years. That averages to roughly $60 billion per year, though the actual profile is likely front-loaded to match Stargate's deployment timeline. Alongside it, OpenAI has floor-level spending built out in specific facility programs: Vice President of Compute Strategy Sachin Katti has stated that OpenAI's Georgia data-centre project alone will cost more than $30 billion when fully developed. These are not forecasts. They are signed or committed spend, the concrete floor beneath the higher aggregate numbers below.

5.2 Stargate: $500 billion over four years with SoftBank, OpenAI, Oracle, and MGX

The Stargate Project, announced by OpenAI in January 2025, is a $500 billion joint venture to build AI infrastructure for OpenAI in the United States over four years. The initial equity funders are SoftBank, OpenAI, Oracle, and MGX. SoftBank holds financial responsibility, and OpenAI holds operational responsibility, a role division that's often mis-reported. Masayoshi Son is chairman. First deployment: $100 billion immediately, with the buildout starting in Abilene, Texas, and additional sites under evaluation across the country. Key initial technology partners include Arm, Microsoft, Nvidia, Oracle, and OpenAI. JPMorgan acted as lead-left, sole underwriter, and sole structuring agent on both financing transactions for Stargate's Abilene campus. Stargate is a separate company, not a line item on OpenAI's balance sheet; it's the vehicle that deploys the $500B on behalf of OpenAI and its partners.

5.3 $280B: the FT's projected 2030 burn

The Financial Times, cited by Reuters on September 18, 2026, reported that OpenAI expects to burn through almost $280 billion in cash between 2026 and 2030, citing a company presentation. The $280B is cumulative outflow across training, compute, and R&D: a company-level projection, not a capex line and not the same as Stargate's $500B (which includes partner capital). The projection sits alongside OpenAI's stated target of 10 gigawatts of data-centre capacity by the end of 2029.

5.4 $750B: the TechCrunch running total across all announced deals

TechCrunch's July 2026 tally put OpenAI's spending across all publicly announced compute, chip, cloud, and infrastructure agreements at roughly $750 billion. It's the broadest possible framing: the sum of signed deals across Oracle, Microsoft, Nvidia, Broadcom, Google Cloud, AMD, and others, over their full contractual duration. Some of those deals overlap. The $300 billion Oracle contract runs concurrently with the $500 billion Stargate project, and both include Nvidia GPU purchases. TechCrunch itself flags the double-counting risk. The $750B is the ceiling framing for OpenAI's total financial exposure; the $300B Oracle contract is the floor. The other three numbers (Stargate $500B, projected burn $280B, running total $750B) sit between them, each measuring a different scope.

5.5 The wheresyoured.at counterframe: is any of this actually payable?

Ed Zitron's wheresyoured.at analysis titled "OpenAI $400bn" provides the skeptical counter-framing. Zitron argues that OpenAI's commitments significantly outstrip any credible path to revenue that would cover them. His breakdown of near-term needs: $200 billion in the next twelve months to reach 10 gigawatts by 2029; $50 billion in the next six months to build a gigawatt of data-centre capacity for the Broadcom chip partnership; $40 billion for 2026 compute; $500 million for a consumer device. He notes that OpenAI's $391.5 billion funding need is 5.7 times the $67.6 billion in capital expenditures Amazon spent building AWS. This is opinion, clearly labelled, not consensus, but the underlying claim (that the aggregate commitments run ahead of revenue projections) is the central financial-sustainability question the sector has yet to answer publicly.

6. Anthropic: $517 billion in compute agreements, $50 billion in US infrastructure, and Google as banker

6.1 $517 billion in compute agreements in eleven months

Between September 2025 and August 2026, Anthropic signed $517 billion in compute-capacity commitments: a cumulative counterparty total across cloud and chip providers, per The Information and DataCenterDynamics' aggregation. The number is not annual spend and not on-balance-sheet capex; it's the total purchase-commitment value spread over multi-year deal windows. Named components: a $200 billion agreement with Google for TPU access, a $45 billion August 2026 compute deal with Nscale, a $50 billion partnership with AI cloud firm Fluidstack (with data centres being built in Texas and New York), and (separately confirmed via Amazon) Anthropic's commitment to spending more than $100 billion over the next ten years on Amazon Web Services. Since October 2024, per The Information, Anthropic has signed on for a total of 14.8 GW of compute capacity that it will access over the coming years. That capacity total is the physical anchor beneath the $517B dollar figure.

6.2 The $50B American AI infrastructure announcement

In November 2025, Anthropic announced $50 billion in investment in American AI infrastructure over an unspecified multi-year window. This is distinct from the $517B in compute-purchase agreements: the $50B is oriented at owned or co-built infrastructure inside the United States, including data centres and sovereign-compute facilities, framed around US-based capabilities. Company communications emphasise national-resilience framing. Anthropic has not disclosed specific site locations or timelines publicly, but the direction is a dual-track model: leveraging external cloud capacity through the $517B in compute commitments while building strategic US-based assets on top.

6.3 The Akamai deal: $11.6B over seven years

TechCrunch reported in September 2026 that Anthropic will pay Akamai $11.6 billion over seven years in a cloud infrastructure agreement, based on Akamai's securities filing. It's one of the largest single-vendor cloud commitments Anthropic has disclosed, and it moves Akamai from a CDN-and-security vendor into the AI-inference-hosting category. The seven-year duration implies sustained volume assumptions: a bet that Anthropic's inference load will grow to justify multi-year committed capacity at Akamai scale.

6.4 AMD-Anthropic: two gigawatts of MI450 chips plus a $5B AMD investment

In July 2026, per Reuters, Anthropic entered a two-way agreement with AMD: Anthropic will buy up to 2 gigawatts of AMD's latest-generation Instinct MI450 chips, starting in the first half of 2027, and AMD will invest up to $5 billion in Anthropic, with the investment tied to specific deployment milestones. It's a diversification move for Anthropic away from Nvidia dependence, and it's simultaneously the highest-profile customer win AMD has landed against Nvidia in the AI-accelerator market. The 2 GW is a substantial compute commitment, though exact chip counts and deployment timelines aren't specified. What the deal structure suggests is the shape of future chip-supplier relationships in AI: not simple procurement, but bundled equity investment plus multi-year committed capacity.

7. Nvidia: supplier and spender, 68% of chip designs and a $105B stake in OpenAI's infrastructure

7.1 AI chip market share by designer

Nvidia's dominance is real, but it's not what most non-specialist coverage suggests. Per Epoch AI's Q4 2025 chip-market-share-by-designer analysis, Nvidia held 68% of AI chip supply measured in H100-equivalents of AI computing power sold or shipped in the quarter. Google, primarily through its custom TPU line, held 19%: a substantial second position most industry commentary underestimates. Amazon and AMD each held 5%. Huawei held 3%. Cambricon held less than 1%. Total AI computing capacity across all designers has grown roughly 3.3× per year since 2022. That growth number is at least as important as the share number: even at 19% market share, Google's absolute AI computing capacity is expanding rapidly, and Nvidia's share can shrink slightly while its capacity output still rises meaningfully. Designer share is not revenue share and not installed share; it's a measure of design origination for chips sold or shipped that quarter.

7.2 Earnings: what August 2026 disclosed

Nvidia's August 2026 earnings confirmed the company's position as the linchpin of the AI-infrastructure economy. Second-quarter revenue more than doubled to $96.22 billion, beating analyst estimates of $92.17 billion. The company forecast third-quarter revenue of $108 billion, plus or minus two percent, versus analysts' average estimate of $104.19 billion, per data compiled by LSEG. CFO Colette Kress cited the Vera Rubin platform's early customer shipments; Vera Rubin is expected to contribute about a fifth of total data-centre revenue in the current quarter ending in October. Separately, memory producers supplying Nvidia are generating gross margins near 80%, more than twice historical levels, reflecting how AI demand is reordering the component-supplier economics further up the chain.

7.3 The $105B OpenAI backing, and the $500B Nvidia-Wall Street fund

Nvidia's role in 2026 goes beyond selling chips. Per Yahoo Finance, Nvidia will back OpenAI's data-centre buildout with up to $105 billion, and the specific structure, notably, is narrower than most coverage suggests. Nvidia will only pay for portions of lease and power payments, not the full cost of the site or OpenAI's obligations. It's structured participation in specific cost buckets, not a general equity or debt commitment. Alongside the OpenAI backing, Nvidia has set up a $500 billion fund alongside BlackRock, Blackstone, Goldman Sachs, and Apollo Global Management, designed to allow data-centre builders to finance the purchase of Nvidia chips. That fund is arguably the bigger structural move; it's Nvidia acting as a market-making financier for the industry's next generation of chip fleets, not just an equity investor in one customer. Nvidia has also been active elsewhere: a $5 billion investment in Intel giving roughly 4% ownership after new shares are issued, plus stakes in Safe Superintelligence and SK Hynix. The chip supplier is becoming a diversified capital allocator across the AI supply chain.

8. CoreWeave and the neo-clouds: renting your GPUs from a company that borrowed to buy them

8.1 The $35–39B 2026 capex plan

CoreWeave's 2026 capital-expenditure plan projects $35–39 billion in spending, up from a $31–35 billion range in 2025, per Zacks Equity Research. For context, competitor Nebius Group has forecast $20–25 billion in 2026 capex, having spent nearly $5.7 billion in Q2 2026 alone on GPUs and data-centre expansion. Nebius Group is also providing Microsoft with GPU infrastructure capacity in a $17.4 billion deal over five years. The neo-clouds (CoreWeave, Nebius, and their peers) are collectively expected to exit 2026 with more than eight gigawatts of Nvidia GPU capacity, a substantial jump from the year before. CoreWeave's capex-to-revenue ratio is among the highest in the industry: this is a business model built on capacity dominance, not near-term profitability. Interest expense alone hit $640 million recently, up from $267 million a year earlier: a direct measure of how quickly the debt load is scaling.

8.2 How CoreWeave finances the fleet

CoreWeave's financing model is the poster child for chip-backed debt. GPUs serve as collateral for loans that fund the purchase of additional GPUs, with future customer revenue underwriting the debt service. The Wall Street Journal has covered the mechanics of the model in depth: CoreWeave's ability to scale hinges on maintaining strong demand from AI-native customers and stable chip resale values. Unlike hyperscalers with self-funding cash reserves, CoreWeave's business depends on continuous access to credit markets. The specific risk is asymmetric: if GPU resale values fall faster than the debt amortises, the collateral chain shortens, potentially triggering margin calls or refinancing pressure. So far, the model has scaled successfully; the question the market is watching is whether it holds through a demand slowdown or an accelerated chip-generation cycle.

9. How it's actually being financed: chip-backed leases, private credit, and the debt curve

9.1 JPMorgan's framing: AI financing is the biggest secular theme in a generation

JPMorgan's global co-head of Investment Grade Finance, John Servidea, has called AI financing "the biggest secular theme in our professional lifetimes." JPMorgan acted as lead-left, sole underwriter, and sole structuring agent on both financing tranches for OpenAI's Stargate Abilene, Texas campus: the industry's flagship AI-infrastructure financing structure. Behind the framing is a market-structure shift. Capital-markets appetite for long-dated AI-infrastructure exposure is compressing spreads, extending tenors, and enabling structural innovations (asset-backed tranching, embedded performance covenants tied to chip deployment schedules) that would have been niche two years ago. What was speculative risk a decade ago is now institutionalised as core investment strategy. The banker's diagnosis, from the bank underwriting the biggest deals, is that this is early-cycle rather than late.

9.2 Chip-backed leases as the new asset class

The 2026 financing innovation is leases structured against chip fleets and data-centre hardware rather than long-lived building shells. The economic logic: hardware amortises in three to five years while data-centre buildings last twenty to thirty. Financing the two separately allows payment tenors to match actual asset lives, unlocking access to private credit markets that had previously rejected long-term data-centre bets. Meta's $27 billion Blue Owl Capital financing for the Louisiana site (introduced in section 4.3) is one canonical example. CoreWeave's model (introduced in section 8) is another. The broader move is away from balance-sheet-heavy capex and toward asset-light, performance-linked financing: a structural shift that has attracted the largest private-credit shops, including Blue Owl, Apollo, Blackstone, and BlackRock, into the data-centre asset class at unprecedented scale.

9.3 The $1.09 trillion uncommenced-lease liability, revisited

Returning to the $1.09 trillion in uncommenced lease commitments introduced in section 1.4: the concentration matters. Meta's $278.99 billion is the largest disclosed single-company figure, followed by Amazon's $137.21 billion (though Amazon's figure includes warehouses, offices, aircraft, and vehicles, not just data centres, so it's not directly comparable). Alphabet disclosed $85.2 billion. The most-analysed number is Oracle's $260 billion. S&P Global Ratings has incorporated it into its adjusted-debt forecast, and S&P expects Oracle's leverage to run around 4.4× in fiscal 2027. Oracle has explicitly warned in disclosures that the duration, renewal terms, and pricing of its data-centre leases may not align with customer contracts, a mismatch that leaves Oracle exposed if customers don't renew or can't perform. That specific misalignment risk is what analysts and credit-rating agencies are watching. It is the first meaningful stress point publicly disclosed inside the 2026 AI-infrastructure financing structure.

10. The suppliers behind the suppliers: power, cooling, and the rest of the ecosystem

10.1 The $1.3 trillion "energizers" bucket

McKinsey's split of the $5.2 trillion AI-workload total assigns $1.3 trillion, or 25%, to "energizers": companies delivering power generation, transmission, cooling, and electrical equipment through 2030. That number is roughly 4.5× the combined 2026 capex of Google ($205B) and Microsoft ($80B). The power-and-cooling layer isn't a rounding error inside the AI story. It's the largest bucket of AI-infrastructure capex after chips themselves, and much of it flows to companies whose stock tickers most AI-story readers wouldn't recognise.

10.2 The Reuters power-and-cooling piece: who's actually winning

Reuters' September 2026 analysis, "Not just Nvidia: these power and cooling firms are riding trillion-dollar data-centre demand," profiles the second-tier of the supply chain: companies that operate outside the AI headline cycle but capture meaningful share of the buildout. The named beneficiaries are heavily concentrated in Asia. China's Jinpan saw its shares surge 118% during 2025, and Hainan Jinpan Smart Technology reported first-half 2026 new data-centre orders more than quadrupled from a year earlier, with related backlog nearly tripling, driven particularly by AI infrastructure projects in North America. Taiwan's Delta Electronics, a major power-infrastructure supplier and part of Nvidia's supplier ecosystem, has seen shares rise more than 90% year-to-date; Chairman Ping Cheng said in July that even if revenue increases, gross margins will probably remain at current elevated levels. Delta is expanding production footprint across Thailand and the US, and reports that demand for AI power, cooling, and data-centre infrastructure "remains a growth engine." Other Nvidia-ecosystem suppliers profiled include Asia Vital Components and Auras Technology, both in the thermal-management category. Solid-state transformers (SSTs) are emerging as a next-generation technology in the space; UBS forecasts that Chinese companies will gain share in the SST market on technological expertise and cost advantages. Taiwan's Delta reports that a small data centre is already deploying its SSTs. Behind the specific company names is a demand-side dynamic that structures the whole tier: hyperscalers want facilities delivered within six months, and equipment makers with order backlogs stretching more than three years (some booked out to 2030) hold pricing power that hyperscaler capex programs, however large, can't dislodge in the near term.

10.3 What it all adds up to

By 2026, AI infrastructure spending has crossed $1 trillion by Goldman Sachs' measure and appears headed toward $5–7 trillion in cumulative outlays by 2030 by McKinsey's. The total is not a single program but a web of overlapping commitments: hyperscaler capex on the order of $600B–$800B annually, AI-native compute-purchase agreements running into the hundreds of billions each for OpenAI and Anthropic, chip-backed private-credit lease structures pulling forward tens of billions per single site, and roughly $1.09 trillion in uncommenced lease commitments already locked in but not yet appearing on balance sheets. Google, Microsoft, Meta, OpenAI, Anthropic, Nvidia, and CoreWeave are each executing distinct programs that overlap and double-count in the headline numbers. Beneath the surface, a secondary layer of power-and-cooling suppliers captures roughly a quarter of the total spend, with the growth concentrated in Asian equipment makers rather than the traditional Western industrial names. The financing mix is shifting from operating cash to chip-backed leases, private credit, and increasingly, direct capital participation by chip suppliers like Nvidia through funds structured with the biggest asset managers on Wall Street. The industry's true financial footprint extends well beyond server racks and GPU arrays. It runs through grids, transmission lines, cooling towers, transformer factories, and the balance sheets of a rapidly expanding roster of private-credit vehicles that now treat AI infrastructure as their primary asset class.