AI CapEx 2026: Scale, Structure, and Return Paths

Section 01Scale of the Build-Out

The four largest hyperscalers, Amazon, Alphabet, Microsoft, and Meta, have guided to roughly $720 to $765 billion in combined capital expenditure for calendar 2026. Adding Oracle, CoreWeave, and SpaceX pushes the observable U.S. AI-related CapEx total well above $850 billion for the year.

Sequential quarterly increases remain large. In the most recent reported quarter (ended June 2026), Amazon spent $53.1 billion on cash CapEx, Alphabet $44.9 billion, Microsoft $41.0 billion, Meta $31.1 billion, and SpaceX $18.4 billion, of which $15.8 billion was AI infrastructure.

Section 02Four Economic Models

Four practical categories organize the different return paths and risk profiles. They are not rigid classifications.

SpaceX dual-role note. SpaceX is best classified as a Vertical Integrator for its core internal demand (its AI business, formerly xAI, plus physical systems following the 2026 SpaceX–xAI merger). It has, however, become a material external compute supplier. Its agreement with Anthropic, disclosed in SpaceX’s May 2026 S-1 IPO filing, is worth $45 billion in total, structured at $1.25 billion per month through May 2029, with either party able to exit on 90 days’ notice. This dual role, internal demand plus external leasing, is material and should be tracked separately from pure vertical spend.

Section 03Methodology Notes

• Primary source: company-reported CapEx, management guidance, and earnings call transcripts.
• Ranges are preserved where companies give them; fiscal-year figures are flagged separately from calendar-year figures.
• Cross-year and cross-company comparisons use calendar-year framing where possible.
• Free-cash-flow coverage of CapEx is shown as operating cash flow divided by CapEx for the same period.
• No forward valuation multiples or stock recommendations are offered.

Section 04Company Deep-Dives

Amazon

Largest absolute spender. Guidance raised to approximately $220 billion for 2026 on the July 30 earnings call, up from $200 billion, driven by higher memory costs. CEO Andy Jassy told investors the company will still not have enough capacity to meet demand in 2026, and expects that to hold into 2027. AWS revenue grew 37 percent in Q2, the fastest pace in eighteen quarters, with an AWS backlog of $496 billion. Trailing-twelve-month free cash flow moved to an outflow of $7.6 billion, driven by a $66.1 billion year-over-year increase in property and equipment purchases.

Alphabet

Guidance raised to $195–205 billion after Q2 earnings on July 22. Q2 CapEx alone reached $44.9 billion. Google Cloud backlog reached $514 billion, roughly double the figure reported a year earlier. Free cash flow turned negative in Q2, the first negative quarter in roughly two decades, an outflow of approximately $5.9 billion.

Microsoft

Calendar 2026 CapEx in the $175–190 billion range. Azure growth continues at a high rate, with management stating Azure demand exceeds supply. Microsoft has kept free cash flow positive while several peers turned negative, though the margin of safety is narrower than in prior years.

Meta

Guided to $130–145 billion for 2026, including finance-lease principal, narrowed from a wider prior range. Q2 CapEx was $31.1 billion against $31.9 billion of operating cash flow, consuming nearly all of it. Free cash flow fell from $8.5 billion to $784 million year over year, a decline of more than 90 percent.

Oracle

Remaining performance obligations reached a record $638 billion as of May 31, 2026, up 363 percent year over year and up $85 billion in the fourth fiscal quarter alone. Most of the increase reflects large-scale AI contracts, including arrangements where the customer prepays for or supplies GPU hardware directly, a structure that reduces Oracle’s own capital burden. Bank of America has estimated that more than half of the RPO traces to a single customer, OpenAI. Free cash flow was negative $23.7 billion for fiscal 2026, and Oracle plans to raise approximately $40 billion in additional debt and equity financing in fiscal 2027. S&P Global Ratings downgraded Oracle’s credit rating to BBB- in July 2026, citing the capital intensity of the AI buildout.

CoreWeave

Focused neo-cloud pure-play guided to $31–35 billion of 2026 CapEx. Revenue backlog, which per CoreWeave’s own definition includes remaining performance obligations plus other amounts the company estimates will be recognized under committed customer contracts, reached $99.4 billion at the end of Q1 2026, up nearly 50 percent sequentially. Management has stated that CapEx is tied to already-signed customer contracts, with roughly 36 percent of the backlog expected to convert within 24 months and roughly 75 percent within four years.

Tesla

Guided for more than $25 billion of 2026 CapEx, with management indicating spending will continue to grow for the next two to three years. Portions of the spend support AI training compute, semiconductor efforts, vehicle autonomy, and humanoid robotics. Because the compute is intended primarily for internal products, external market recovery matters less to utilization than it does for the Platform Operators.

SpaceX

SpaceX completed the largest IPO in history in June 2026, raising $85.7 billion, following its earlier 2026 merger with xAI. Its first quarterly report as a public company, released August 4, showed Q2 revenue of $7.81 billion, up 92 percent year over year, against a narrowed net loss of $541 million. Q2 CapEx reached $18.37 billion, more than sixfold higher than a year earlier and well above the $13.22 billion analysts had expected, of which $15.83 billion was AI infrastructure. Q1 CapEx had been $10.1 billion, of which $7.7 billion was AI. The company’s AI segment (which includes its former xAI business, the X platform, and cloud services) generated $2.56 billion in Q2 revenue, up 247 percent year over year, but posted a $1.26 billion operating loss.

SpaceX ended Q2 with $100 billion in cash, cash equivalents, and marketable securities, and reported an AI backlog contributing to a company-wide backlog of $47.5 billion. Musk has stated an internal target of a $100 billion AI annualized revenue run rate by year-end. The stock fell as much as 11 percent after the Q2 report despite the revenue beat, on investor concern over the pace of AI capital spending and the absence of guidance.

Section 052026 CapEx by Quarter and Cash-Flow Impact

The full-year guidance numbers are large, but the cash leaves the building in stages. The second-quarter numbers show continued acceleration rather than a plateau.

Operating cash flow still covers CapEx at Microsoft and Meta on a quarterly basis in the most recent data, while Amazon, Alphabet, and especially SpaceX show materially lower coverage ratios. Full-year arithmetic for Amazon, roughly $97–106 billion combined for the first half against the $220 billion full-year target, implies a second-half run-rate of approximately $114–123 billion, consistent with the guidance within normal quarterly variability.

Section 06Supporting Infrastructure and Peripheral Players

A second layer of companies supplies the storage, servers, networking, power, and cooling systems that turn capital budgets into working data centers. This category sits adjacent to the four economic models above; it is not folded into any single one of them. To keep the list focused, two filters apply: market capitalization above $25 billion, and clear, material exposure to AI infrastructure spend rather than pure software or services.

Storage and Data Management

NetApp provides high-performance enterprise storage systems and software used in AI training and inference environments, helping manage the rapid growth of unstructured data that accompanies large model workloads.

Servers and Systems

Dell Technologies is a major supplier of AI-optimized servers, storage arrays, and full-rack solutions sold to both hyperscalers and large enterprises, with a growing share of revenue tied directly to AI infrastructure deployments.

Networking

Cisco Systems supplies the switching, routing, and data-center networking fabric that connects GPU clusters at scale. Arista Networks focuses more tightly on high-speed Ethernet switches designed for AI and cloud workloads.

Power and Cooling

Vertiv designs power-management and liquid-cooling systems required by high-density AI racks. GE Vernova and Eaton supply power-generation equipment, grid infrastructure, and electrical systems that support the electricity demand of new data centers.

These companies do not set the overall CapEx budget, but they convert a large portion of it into physical capacity. Their order books, backlog growth, and utilization commentary serve as useful secondary signals of how quickly the AI build-out is progressing. Exposure is real but second-order: returns depend on the durability of the primary CapEx cycle rather than on model performance itself.

Section 07Private Frontier Demand

Large multi-year compute commitments from frontier model developers form an important demand bridge for the public infrastructure builders. OpenAI and Anthropic are private and are therefore excluded from any ranking; they are treated solely as demand catalysts.

OpenAI has raised its projected compute spending to approximately $750 billion through 2030. Publicly reported agreements include a roughly $300 billion multi-year contract with Oracle (a primary driver of Oracle’s RPO growth), an AWS commitment totaling approximately $138 billion over eight years, and substantial additional Azure spending, alongside a roughly $250 billion incremental Azure commitment.

Anthropic has announced an up-to-5-gigawatt agreement with Amazon, including nearly 1 GW of new capacity by the end of 2026; a 5 GW agreement with Google and Broadcom beginning in 2027; a strategic partnership with Microsoft and Nvidia including $30 billion of Azure capacity; and a $50 billion U.S. infrastructure investment with Fluidstack. It has separately agreed to use all compute capacity at SpaceX’s Colossus 1 facility, more than 300 megawatts across more than 220,000 Nvidia GPUs, at $1.25 billion per month through May 2029, a deal worth $45 billion in total as disclosed in SpaceX’s IPO filing.

These contracts convert CapEx into contracted revenue for the infrastructure owners and reduce near-term utilization risk. They do not eliminate the longer-term question of whether the underlying models generate sufficient end-user value to sustain the spending cycle.

Section 09Infrastructure Scorecard

A structured worksheet for the reader to fill in as each quarter’s results arrive. Track the six signals below against each company covered in this report.

Absolute CapEx trajectory: actual spend versus most recent guidance, and direction of any revision. Cash-flow coverage: operating cash flow divided by CapEx, same period. RPO / backlog growth: sequential change in remaining performance obligations or contracted backlog. Capacity additions: gigawatts brought online versus management commentary on remaining shortages. External leasing share: especially relevant for SpaceX, given its dual internal/external role. Physical-AI milestones: Tesla autonomy miles, Optimus deployment rates, versus pure cloud utilization (Tesla and SpaceX only).

Section 10Segment Leaders and Forward Indicators

Near-term observable signals matter more than narrative. Retail investors should watch, each quarter:

• Whether CapEx guidance rises, holds, or is cut, and whether companies begin issuing formal multi-quarter guidance where they currently do not (SpaceX being the clearest current gap)
• Cash-flow coverage ratios, and whether any currently positive company (Microsoft, so far) turns negative
• Growth in multi-year contracted cloud and AI services revenue or RPO
• Gigawatts of new capacity brought online versus management commentary on remaining shortages
• For Vertical Integrators: measurable progress in physical systems that can absorb the compute, such as autonomy miles or robot deployment counts
• Any shift in the mix between internal and external demand, particularly at SpaceX

The distinction between software or chatbot AI and physical AI remains important. The former can scale revenue faster once models improve; the latter faces longer hardware, safety, and real-world integration cycles. Both are funded by the same CapEx wave, but their return timelines differ.

Section 11Closing Observation

The absolute scale of 2026 AI CapEx is unprecedented. The more useful questions are structural: which economic models convert the spend into durable cash returns, how quickly payback occurs, and how interdependent the participants have become. The data above is intended as a working map rather than a forecast. Readers can update the tables and scorecard as new quarterly results arrive. No stock recommendations are offered.