The AI infrastructure story in April 2026 is no longer "is the spending big?" — it's "can the grid, the fabs, and the balance sheets keep up?" Hyperscalers are on track to spend close to $700 billion in capex this year. Nvidia just guided to a half-trillion-dollar Blackwell-plus-Rubin pipeline. Anthropic signed a $100 billion ten-year compute deal with AWS. And the IEA is warning that data centers will pull more electricity than all of Japan in 2026. Here's where things actually stand as of late April.

---

TL;DR — six numbers that define AI infrastructure in April 2026

1. Hyperscaler capex is tracking near $700 billion in 2026, up roughly 60% year over year. Amazon at ~$200B, Alphabet at $175–185B, Meta at $115–135B, Microsoft at $120B+, Oracle at $50B. About 75% of that — roughly $450B — is going directly into AI servers, GPUs, and data centers.
2. Nvidia is guiding to ~$500B in combined Blackwell and Rubin revenue through the end of 2026, and ~$1 trillion in orders through 2027. Q4 FY2026 revenue: $68.1B. Full-year FY2026 revenue: $215.9B, with $193.7B from data center. Blackwell is sold out through mid-2026.
3. Anthropic signed a $100B, 10-year AWS compute commitment on April 20, 2026. The deal secures up to 5 GW of capacity, with Claude now training on more than 1 million Trainium2 chips. Amazon is investing up to another $25B in Anthropic on top of the $8B already deployed.
4. Stargate is back on track at ~7 GW of planned capacity and $400B+ of investment over three years. Abilene, Texas is partly online; Michigan, Wisconsin, Wyoming, Pennsylvania, Lordstown OH, and Milam County TX sites are under construction.
5. The IEA projects data center electricity consumption will hit ~1,100 TWh in 2026 — roughly Japan's entire national electricity use. Data center / SMR offtake agreements have grown from 25 GW in late 2024 to ~45 GW today.
6. EU greenlit up to 5 AI Gigafactories under a €20B InvestAI facility, while UAE, Saudi Arabia, and India accelerate sovereign AI builds. France-UAE alone is funding a $30–50B, 1 GW AI data center project.

The rest of this post is the version with receipts, organized around the five threads that matter most.

---

1. Hyperscaler capex: the $700 billion sprint

The headline number for 2026 is the combined capex of the Big Five US AI infrastructure providers: Microsoft, Alphabet, Amazon, Meta, and Oracle. Estimates from CreditSights, Futurum, and CNBC's reporting on earnings calls converge on a range of $660–700 billion for 2026 — up roughly 60% from 2025's already-historic levels.

The breakdown reported by CNBC and corroborated by multiple sell-side desks:

Company	2026 capex (projected)	Direction
Amazon	~$200B	Most, but not all, for data centers
Alphabet	$175–185B	TPU buildout + cloud
Meta	$115–135B	AI training + serving
Microsoft	$120B+	Azure AI capacity
Oracle	$50B	OCI / Stargate footprint

Roughly 75% of that capex is AI-specific — servers, GPUs, custom silicon, networking, and the data center shells to put them in — rather than traditional cloud refresh.

The cash-flow strain is showing up in the financials. Amazon is projected to turn free-cash-flow negative in 2026. Pivotal Research projects Alphabet's free cash flow to fall almost 90% to ~$8.2B from $73.3B in 2025. Tech megacaps are increasingly funding the buildout with debt and structured financing rather than pure operating cash, which is part of what's driving the bubble debate further down this post.

Stargate: still the biggest single project

OpenAI's Stargate — the joint venture with Oracle and SoftBank announced in January 2025 — went through a public rough patch in mid-2025 with reported disagreements over control. As of February 2026, the partners say it's back on track:

• ~10 GW of total planned capacity underwritten by ~$500B of investment commitments.
• ~7 GW of planned capacity and $400B+ of investments locked in over the next three years across announced sites.
• Abilene, TX flagship: First two buildings operational since September 2025. NVIDIA GB200 racks delivered. Oracle will ultimately deploy more than 450,000 GB200 GPUs at Abilene under a 15-year lease. Remaining six buildings expected to complete by mid-2026.
• New US sites under construction: Michigan (1 GW, $7B, Saline Township), Wisconsin, Wyoming, Pennsylvania, Lordstown OH (SoftBank-led), and Milam County TX (with SB Energy).

Tom's Hardware and others reported earlier that the JV had not yet hired staff and that no facilities had technically been built under the joint entity itself — but that's largely a corporate-structure issue. The underlying buildout, executed by Oracle and SoftBank Energy, is proceeding.

---

2. Chips: Rubin arrives, custom silicon eats share

Nvidia remains the dominant story, but the most interesting move in April 2026 is how aggressively custom silicon is starting to take share.

Nvidia: Rubin in volume H2 2026, $1T orders through 2027

Nvidia formally launched the Vera Rubin platform at CES 2026 and reiterated the timeline at GTC in March 2026: - All Rubin-family chips are in the fab at TSMC. - Volume production: H2 2026. - HBM4 memory. - Rubin R100 GPU reported to feature ~336 billion transistors — roughly a 5x leap over Blackwell on certain workloads.

The financial picture from Q4 FY2026 (reported February 2026) was the largest AI revenue print in history: - Q4 FY2026 revenue: $68.1B. - Full-year FY2026 revenue: $215.9B. - Data center revenue: $193.7B for the year. - Blackwell backlog: ~3.6 million units, sold out through mid-2026.

Jensen Huang told CNBC at GTC that he sees ~$500B in combined Blackwell + Rubin revenue through end of 2026 and ~$1 trillion in cumulative orders through 2027. Supply, not demand, is the binding constraint.

Custom ASICs are growing faster than GPUs

The most underrated chart in AI infrastructure right now: custom ASIC shipments from cloud providers are projected to grow ~44.6% in 2026, vs. ~16.1% for GPUs. That doesn't mean Nvidia loses — the absolute base is much larger — but the silicon mix is shifting.

The big custom-chip stories:

Google TPU. On April 22, 2026, Google unveiled new training and inference TPUs that it says deliver 2.8x the performance of Ironwood (its 7th-gen TPU announced November 2025) at the same price, with 80% better performance on the inference variant. Anthropic also publicly committed to training Claude on up to 1 million TPUs. Broadcom confirmed a long-term agreement with Google to develop and supply future-generation TPU silicon.

AWS Trainium. Project Rainier is now fully operational with ~500,000 Trainium2 chips across multiple US data centers — one of the largest AI training clusters ever built. Anthropic's $100B AWS commitment (more on this below) explicitly covers Trainium2 capacity coming online in H1 2026 and Trainium3 capacity by end of 2026, scaling to ~5 GW total.

Cerebras. Closed a $1.1B Series G to scale its wafer-scale CS-3 systems into AI supercomputers without distributed training overhead.

Groq. Reporting 276–300 tokens/sec on Llama 70B, up to 1,665 tokens/sec with speculative decoding, at $0.79 per million output tokens for Llama 3.3 70B. Speed-as-a-service is the pitch.

Broadcom. The unsung beneficiary. Custom-silicon partnerships with both Google (TPU) and Anthropic (~3.5 GW of TPU-based capacity starting in 2027) put Broadcom at the center of the non-Nvidia compute stack.

---

3. Power: the constraint that's actually binding

If 2024 was the year of "GPU scarcity" and 2025 was the year of "data center concrete and steel," 2026 is the year of power. The IEA, Uptime Institute, and Morgan Stanley all flagged it in Q1 reports.

Key numbers: - IEA projects global data center electricity consumption at ~1,100 TWh in 2026 — roughly equal to Japan's entire annual consumption. - Uptime Institute projects total global AI-specific data center load will hit ~10 GW by end of 2026. - Conditional offtake agreements between data center operators and small modular reactor (SMR) projects have grown from ~25 GW at end of 2024 to ~45 GW today.

The nuclear pivot

January 9, 2026 was the marquee day for nuclear-AI deals: Meta signed agreements with three nuclear companies for 6+ GW of power. That came on top of Meta's earlier 20-year, 1.1 GW Constellation Energy deal in Illinois beginning in 2027. Microsoft, Amazon, and Google have all signed nuclear PPAs in the past 18 months — the question is no longer "is nuclear back?" but "how fast can SMRs actually get permitted and built?"

"Bring your own power"

Because utilities can't deliver interconnection on hyperscaler timelines, the new pattern is on-site generation. The IEA noted in its 2026 update that data center developers are advancing a large pipeline of projects with on-site natural gas generation in the US, plus microgrids, batteries, and hybrid systems. The phrase showing up in operator earnings calls: "energy islands."

This has second-order consequences: - It shifts the bottleneck from grid interconnect queues to gas turbine and SMR supply chains (themselves multi-year lead times). - It puts hyperscalers in the unusual position of becoming meaningful electricity producers. - It makes data center siting decisions function as de-facto industrial policy at the state and federal level.

---

4. Inference and serving: the layer where economics get real

Training gets the headlines, but inference is where the unit economics actually have to work — and 2026 is the year inference infrastructure became its own multi-billion-dollar tier.

The Anthropic / AWS deal: $100B over 10 years

April 20, 2026 was the biggest single inference-relevant deal of the year so far: - Amazon committed up to another $25B in equity into Anthropic, on top of the $8B previously invested. - Anthropic committed $100B+ in AWS spend over 10 years, primarily on Trainium silicon. - Up to 5 GW of compute capacity secured for training and serving Claude. - ~1 GW of combined Trainium2 + Trainium3 capacity coming online by end of 2026.

The financial engineering here is worth flagging — Amazon is investing equity into Anthropic that flows back as AWS revenue. That's not unique to this deal (see the circular-financing section below), but the scale is.

The independent inference layer

Below the hyperscalers, a tier of inference-focused providers is consolidating: - Together AI and Fireworks AI compete on broad model support, fine-tuning, and price-performance. - Groq competes purely on latency, leveraging its custom LPU. - inference.net, Replicate, and others fill the long tail.

Across these providers, Llama 4 Scout and Maverick API pricing has compressed to $0.07–$0.90 per million input tokens as of April 2026 — a meaningful drop from Q4 2025 pricing. That compression is one of the most important — and most under-reported — data points for the bubble debate. If inference cost-per-token keeps falling at this rate, the revenue ramp required to justify $700B of capex looks more achievable; if not, the math gets uncomfortable.

Ramp's customer-spend telemetry shows AI infrastructure businesses pulled ~$260M in Q4 2025 spend from Ramp customers — about 60% of what those same customers spent directly on closed-source foundation model APIs. Infrastructure is growing, but model APIs are still the bigger line item.

---

5. Sovereign AI: the rest of the world stops waiting

Q1 2026 was, in retrospect, the quarter sovereign AI infrastructure became real policy rather than press releases.

EU AI Gigafactories

On January 16, 2026, the Council of the EU adopted an amendment to the EuroHPC Joint Undertaking regulation that explicitly authorizes EuroHPC to fund, develop, and operate AI Gigafactories — large-scale facilities for training trillion-parameter models. The InvestAI Facility provides a €20 billion fund to support up to 5 AI Gigafactories across member states via public-private partnerships.

This is the EU's industrial-policy answer to US hyperscaler dominance. It is separate from the EU AI Act regulatory file but increasingly entangled with it — the same Commission writing the rules is also writing the cheques to build EU compute.

Middle East: the $100B+ buildout

• UAE-France partnership: A $30–50B, 1 GW AI data center project in France, with the UAE as a strategic partner. The agreement introduced the concept of "virtual data embassies" — a sovereignty workaround that allows joint operation while preserving data jurisdictional control.
• Saudi Arabia HUMAIN: Launched May 2025 under the Public Investment Fund, planning 11 data centers totaling 2,200 MW powered by hundreds of thousands of Nvidia GPUs.
• Saudi-Nvidia sovereign AI factory: Up to 5,000 Blackwell GPUs for SDAIA's sovereign deployment.

India, Japan, UK, Indonesia

All four launched sovereign AI compute initiatives in Q1 2026. The pattern: a national fund, a designated cloud or telco partner, an Nvidia GPU allocation, and a sovereign-data clause. Specific scale varies wildly — the UK and Japan are pursuing tens of thousands of GPUs; India's program is structured differently, with multiple regional clusters rather than a single mega-facility.

---

The bubble debate: don't pretend it's settled

You cannot write an honest April 2026 AI infrastructure update without addressing this. Both sides have evidence.

Bear case (concrete)

• Circular financing. Bloomberg's March 2026 graphics package mapped how Microsoft, OpenAI, and Nvidia keep paying each other. Most cited example: Nvidia invests $100B in OpenAI; OpenAI's CFO acknowledges most of that money flows back to Nvidia for GPU purchases. Nvidia holds a $3B stake in CoreWeave; CoreWeave has spent ~$7.5B buying Nvidia GPUs. Anthropic-AWS is structurally similar, just larger.
• Revenue gap. OpenAI annualized revenue: ~$25B. Anthropic annualized revenue: ~$19B. Combined frontier-lab revenue is roughly 6% the size of one year of hyperscaler capex. That gap has to close — fast — or the spending gets re-rated.
• Capex as % of GDP. Multiple analyses note that 2026 hyperscaler AI capex, measured against US GDP, exceeds the Apollo program, the interstate highway system, and the railroads — only the Louisiana Purchase clears it.
• Opaque debt. AI infrastructure is increasingly funded via SPVs, private credit, asset-backed securities, and credit default swaps. Vanderbilt's "After the AI Crash" paper (March 2026) and Time's "Prepare for an AI Bubble Now" piece both flag the systemic-risk implications.

Bull case (also concrete)

• Demand exceeds supply. Nvidia Blackwell is sold out through mid-2026. Hyperscaler backlogs are growing, not shrinking. If this were a demand bubble, you'd see inventory build, not stockouts.
• Inference cost is collapsing. Token pricing has compressed roughly 50–80% on commodity model APIs since early 2025. That widens the application surface that's economically viable.
• Custom silicon is real. TPU, Trainium, and the Broadcom/Google and Broadcom/Anthropic partnerships are not vapourware — Project Rainier (~500K Trainium2 chips) is operational.
• Sovereign demand. Even if private-sector AI revenue disappointed, governments are committing tens of billions to sovereign compute as a strategic asset, which puts a floor under hyperscaler and Nvidia order books.

The honest read in April 2026 is that both sets of facts are true simultaneously, and the question is which dynamic dominates over the next 18 months. Anyone telling you it's settled — in either direction — is selling something.

---

What to watch in the next 90 days

A practical list for May–July 2026:

1. Q1 FY2027 hyperscaler earnings (late April / early May). Capex guidance updates from Microsoft, Alphabet, Meta, and Amazon will reset the $700B baseline up or down.
2. Nvidia Q1 FY2027 earnings (late May). First read on Rubin volume production and any Blackwell supply-chain adjustments.
3. Stargate Abilene completion milestones. Buildings 3–8 are scheduled for mid-2026 — slips will be a leading indicator of broader build constraints.
4. EU Gigafactory shortlist. The first round of consortia selections under the InvestAI €20B facility is expected in the summer.
5. US grid interconnect actions. PJM, ERCOT, and MISO queues are the binding US constraint. Any FERC or state-PUC actions to fast-track AI-related interconnection will materially change capex pacing.
6. Inference token-price floor. If commodity inference falls another 30–50% in H1 2026, the bull case strengthens. If it stalls, the bear case does.

---

Bottom line

The AI infrastructure story in April 2026 is no longer a single question. It's four questions stacked on top of each other:

• Can hyperscalers actually spend $700B in a year? (Answer: physically yes, if they get the power and the chips. Financially, the cash-flow strain is real.)
• Can Nvidia deliver Rubin on the announced timeline? (H2 2026 volume — track GTC and earnings updates closely.)
• Can the grid keep up? (Probably not without on-site generation and SMR breakthroughs.)
• Can AI revenue catch up to AI capex? (The most important question, and the most genuinely uncertain.)

If you're operating in this market — building products, evaluating vendors, or making capital-allocation calls — the right posture isn't "this is a bubble" or "this is the next industrial revolution." It's: plan for a world where both stories are partly right. Procurement should still treat compute as scarce. Finance should still stress-test for a re-rating. Strategy should still assume Nvidia's lead is durable but not monopolistic.

The deals are real. The chips are shipping. The power is the bottleneck. The financials are stretched. All four are true at the same time — and that's the actual news.

---

Sources and further reading

Hyperscaler capex - CNBC: Tech AI spending approaches $700 billion in 2026, cash taking big hit - Futurum: AI Capex 2026 — The $690B Infrastructure Sprint - IEEE ComSoc: Hyperscaler capex > $600bn in 2026 - CreditSights: Technology — Hyperscaler Capex 2026 Estimates - Epoch AI: Hyperscaler capex has quadrupled since GPT-4's release - Network World: Hyperscaler backlogs show growing demand for AI infrastructure

Stargate - OpenAI: Five new Stargate sites with Oracle and SoftBank - OpenAI: Stargate advances with 4.5 GW partnership with Oracle - DataCenterDynamics: OpenAI announces five more US Stargate data centers - Tom's Hardware: Stargate AI data centers reportedly delayed by partner squabbles - IntuitionLabs: Stargate Project — Guide to the AI Infrastructure

Nvidia / Rubin / Blackwell - CNBC: Nvidia GTC 2026 — Huang sees $1T in orders for Blackwell and Vera Rubin through 2027 - Yahoo Finance: Nvidia launches Vera Rubin at CES 2026 - DigiTimes: Nvidia sees $500B Blackwell-Rubin pipeline through 2026 - Seeking Alpha: Nvidia targets $65B Q4 revenue and $0.5T Blackwell + Rubin sales - TweakTown: Nvidia confirms next-gen Rubin AI GPUs with HBM4, volume production H2 2026 - Tech-Insider: Nvidia Vera Rubin Platform — 336B Transistors and 5x Blackwell Leap

Custom silicon - CNBC: Google unveils chips for AI training and inference in latest shot at Nvidia (April 2026) - TechWire Asia: Broadcom's custom AI chips deal with Google and Anthropic explained - StorageReview: From CPUs to TPUs — The Custom Silicon Revolution - Introl: AI Accelerators Beyond GPUs — TPU, Trainium, Gaudi, Cerebras

Anthropic / AWS / Project Rainier - Anthropic: Anthropic and Amazon expand collaboration for up to 5 GW of new compute - CNBC: Amazon to invest up to another $25 billion in Anthropic as part of AI infrastructure deal - TechCrunch: Anthropic takes $5B from Amazon and pledges $100B in cloud spending - About Amazon: AWS activates Project Rainier — one of the world's largest AI compute clusters - DataCenterDynamics: AWS activates Project Rainier cluster of nearly 500,000 Trainium2 chips

Power / energy - IEA: Data centre electricity use surged in 2025 — bottlenecks driving solutions - TechCrunch: Meta signs deals with three nuclear companies for 6+ GW of power - Morgan Stanley: Energy markets race to solve the AI power bottleneck - Belfer Center: AI, Data Centers, and the U.S. Electric Grid — A Watershed Moment - DataCenterKnowledge: How realistic is nuclear power for AI data centers? - Global Data Center Hub: Q1 2026 — The Quarter AI Infrastructure Became Energy-Constrained

Inference - Infrabase: AI Inference API Providers Compared (2026) - Inference.net: LLM API Pricing Comparison 2026 - Ramp: What Ramp data reveals about the fast-growing AI infrastructure market

Sovereign AI / Gigafactories - European Commission: AI Factories — Shaping Europe's Digital Future - BABL AI: EU greenlights AI Gigafactories in push for strategic compute - European Economics: InvestAI initiative and the AI Gigafactories call - Introl: Middle East AI Revolution — UAE and Saudi Arabia's $100B+ Infrastructure Plans - Raise Summit: Sovereign AI — Why Nations Are Treating Compute as Critical Infrastructure - WEF: How shared infrastructure can enable sovereign AI

Bubble debate - Bloomberg: AI Circular Deals — How Microsoft, OpenAI and Nvidia Keep Paying Each Other - Time: We Must Prepare For an AI Bubble Now - Vanderbilt: After the AI Crash (March 2026) - Man Group: The AI Bubble — Hidden Risks and Opportunities - StartupHub.ai: The Circular Funding Deals Driving the AI Boom