AI Capex 2026: The $700bn Build-Out Is a Subsidy for the Application Layer

AI capex 2026 is a $700bn line item. Microsoft, Alphabet, Amazon and Meta plan roughly that much infrastructure spending this year, more than double what the entire global telecom industry spends annually, per Benedict Evans' May 2026 presentation. The strange part is who the spending actually subsidizes.

The evidence in the same deck points one layer up. Frontier models are converging on benchmarks, inference efficiency improves 50-100x per year, and the labs cannot build every application themselves. Whoever turns commodity intelligence into deployed workflow captures the durable margin.

That reading matters most in markets that never had capital to burn. Avante Ventures builds AI-native companies in Brazil and Latin America precisely because someone else is paying for the infrastructure.

The $700bn build-out, in numbers

The big four plan roughly $700bn of capex in 2026 per company guidance including capital leases, compiled by Benedict Evans. Global telecoms spend about $300bn a year. Oil and gas runs about $1tr. AI infrastructure now outspends every telecom operator on earth combined, twice over.

These were asset-light businesses a few years ago. Capex to sales for 2026 estimates runs near 55% at Meta, 54% at Microsoft, 44% at Alphabet and 26% at Amazon. Free cashflow machines have turned themselves into industrial builders.

The buyers admit the logic is defensive. Sundar Pichai says the risk of under-investing is significantly greater than the risk of over-investing. Zuckerberg's stated worst case is having just pre-built for a couple of years. Nobody wants to be the one who stopped digging.

• US data centre construction, excluding the compute itself, is overtaking US office construction, each near a $50bn seasonally adjusted annual rate in early 2026, per US Census data cited by Evans.
• Semiconductor capex from TSMC, Intel and Samsung is guided to about $145bn in 2026. Nvidia books $65-70bn in quarterly revenue while Intel sits flat at $13-14bn.
• Financing is moving off balance sheet: Meta's $27bn Hyperion JV with Blue Owl Capital, Oracle bonds trading like junk on data-center-completion fears, and OpenAI discussing 30GW+ of capacity deals at roughly $20bn per gigawatt.

Models are converging into commodities

Frontier models from OpenAI, Anthropic, Google, Meta and the Chinese labs now cluster tightly on aggregate benchmark scores, per ArtificialAnalysis data in Evans' deck. A new frontier ships every 6-9 months and resets the race. Inference efficiency improves 50-100x per year, so yesterday's expensive capability becomes today's cheap API call.

The revenue is real but small against the spend. OpenAI's net run-rate approaches $2bn a month and Anthropic's gross run-rate sits near $3.5bn a month by mid-2026, per company figures and press reports. Set that against a $700bn build and the equilibrium is missing.

Evans' structural point is the uncomfortable one. Models so far look like commodities: capital-intensive, no network effects, possibly low margin. Classic software is the exact opposite. Sam Altman said it plainly, describing a future where intelligence is a utility like electricity or water, bought on a meter. Utilities are essential. Utilities are rarely where the margin lives.

The telco lesson: infra rarely captures the value

Mobile operators built a trillion dollar industry and watched the value get captured by other people. Global mobile data traffic grew about 30x from 2010 to 2025 while the MSCI global telecom index went roughly sideways, per Evans. Uber, Spotify and WhatsApp were built on those networks. The networks got the bill.

The AI version of that story is already visible. Chat is a poor interface for most real work, so general use needs purpose-built applications, and the labs cannot build or generate them all. If the model layer is infrastructure, innovation and margin move to whoever owns the use case.

Recorded music shows the other edge of the same blade. Industry revenue fell from about $40bn in 2000 to $18bn in 2015 because the internet removed the physical cost base that was the industry's actual moat. When a layer commoditizes, whoever treated that layer as their moat loses it. Every operator should ask which of their costs is quietly playing that role.

A mile wide, an inch deep: the deployment gap

ChatGPT passed 900m weekly active users by early 2026 and only about 5% of them pay, per OpenAI figures in Evans' deck. Depth is thinner than the headline: even the top 20% of users sent fewer than roughly 2,000 messages in all of 2025. For at least 80% of users this is not yet a daily tool.

Enterprises are stuck one step earlier. Bain adoption data shows most functions at 40-70% pilot while production deployment lags far behind, around 40% for software development and 3% for legal. US CFOs told the Atlanta Fed they realized about 1.8% of productivity impact in 2025. Nearly everyone has a pilot. Few have a line on the P&L.

Coding is the exception that proves the point. It is the one use case where the workflow, the evaluation criteria and the buyer were already aligned, and it maxed out Uber's full-year AI budget within months of 2026 starting, per The Information. Every other vertical is waiting for someone to build its equivalent.

Where the profit pool actually forms

The durable assets sit exactly where a model API stops: proprietary workflow data, domain-specific evals, distribution into an industry, and trust. Accenture books about $2.2bn of generative AI work per quarter because enterprises pay for deployment, not for tokens. YC batches went from about 15% AI startups before 2020 to 85-90% by 2025. The unbundling has started.

The mechanism is data network effects in vertical AI. Each deployed workflow generates data no competitor can rent, the data improves the product, and the better product wins more workflow. Avante runs this loop deliberately across its portfolio through the copilot to data to fund flywheel: build an AI copilot to generate proprietary data, then use that data to raise and deploy capital.

Zuckerberg's line about one or two people shipping in a week what used to take dozens of people months is true and beside the point. Writing code was never the hard part. Knowing what the code should do, and owning the workflow it lives in, is the scarce asset.

The honest risks for application builders

Nobody knows whether this capex cycle ends in overcapacity or scarcity. Evans says exactly that, and pretending otherwise would be marketing. Three risks deserve a straight answer.

The Philippines is the cautionary case for services economies. IT-BPM outsourcing employs about 1.9m people, near 8% of GDP, built on skill and income arbitrage that LLMs directly attack. Services work that is a task rather than a job gets automated. The open question is who owns the automation when it happens.

• The labs are climbing the stack themselves. OpenAI runs Frontier Alliances with BCG, McKinsey, Accenture and Capgemini, and announced a $10bn joint venture with PE firms in May 2026. Anthropic signed a $1.5bn JV with Blackstone, Hellman & Friedman and Goldman the same month.
• Cheap intelligence lowers the barrier for your competitors too. A wrapper on a commodity model is a feature waiting to be absorbed. The moat is the proprietary data loop, never the model.
• History does not promise a soft landing. Elevator attendants disappeared entirely after automation. Accountants roughly tripled as a share of US employment while computers ate the mechanics. Uber ate New York taxis, Airbnb only nibbled hotels. It depends.

How Avante builds on top of the capex war

Avante Ventures is a venture studio building AI-native companies in Brazil and Latin America, and the capex war is the studio's tailwind. Services account for roughly 70% of Brazilian GDP, per IBGE, with low software penetration. Commodity intelligence priced like a utility means a Brazilian venture rents the output of a $700bn build for cents. AI infrastructure is now cheap enough to deploy without a Series A.

The build is systematic, not opportunistic. Avante launches 3-4 ventures per year through a six-stage system, Research, Partner, Build, Traction, Revenue, Compound, deploying $500K-1.5M per venture and retaining co-founder economics. The studio benchmark explains the wager: GSSN reports ~50% IRR for studio-built companies versus ~19% for traditional VC. That is the model's benchmark, not Avante's own return, and it is the gap a systematic builder plays for.

Evans closes his deck with the only two honest answers to any AI question: no-one knows, and what happened the last time everything changed. For operators the follow-ups are sharper. Which tasks become free. Was that cost base your moat. What was impossible that is now cheap. The $700bn is someone else's money. The application layer it subsidizes belongs to whoever shows up with the workflow.

Frequently asked questions

How big is AI capex in 2026 and who is spending it?

Roughly $700bn, planned by Microsoft, Alphabet, Amazon and Meta, per company guidance compiled in Benedict Evans' May 2026 presentation. That is more than double the ~$300bn the global telecom industry spends per year, and it is increasingly financed off balance sheet through vehicles like Meta's $27bn Hyperion JV.

Why does the AI capex 2026 boom favor application-layer startups?

Because the spending commoditizes intelligence. Frontier models converge on benchmarks, inference efficiency improves 50-100x per year, and models have no network effects, so the durable margin moves to applications that own proprietary workflow data, domain evals and distribution.

Are AI models becoming commodities?

So far, yes. Benchmark scores from OpenAI, Anthropic, Google, Meta and Chinese labs cluster tightly by 2026 per ArtificialAnalysis, and a new frontier ships every 6-9 months. Models are capital-intensive with no network effects, which is the profile of a commodity, not of classic high-margin software.

What does the AI capex 2026 build-out mean for LATAM startups?

It neutralizes LATAM's historic capital disadvantage. Services account for roughly 70% of Brazilian GDP per IBGE and remain under-digitized, while commodity intelligence priced like a utility means AI infrastructure is now cheap enough to deploy without a Series A. The venture rents the output of a $700bn build for cents.

Is building on top of commodity AI models a defensible moat?

No. Cheap intelligence lowers the barrier for competitors too, so a thin wrapper is a feature waiting to be absorbed. Defensibility comes from the proprietary data loop, which is why Avante runs the copilot to data to fund flywheel: build an AI copilot to generate proprietary data, then use that data to raise and deploy capital.