State of AI
Q3 2026

Let’s start with a number that is almost impossible to put in context: $650 billion. That is what the four biggest technology companies have committed to spending on AI infrastructure in 2026 alone. Not over five years. This year.

Amazon is spending $200 billion. Google is spending $175-185 billion, nearly double what it spent in 2025. Meta has earmarked up to $135 billion. Microsoft is running at a pace that would put it above $145 billion annualised. These are not research budgets or vague strategic commitments. They are contracts already signed, data centres already under construction, GPU orders already placed.

To put it a different way: Goldman Sachs projects $7.6 trillion in cumulative AI capex between 2026 and 2031. The entire GDP of Japan is about $4.2 trillion. The AI buildout, if these projections hold, will consume almost double that over six years.

The obvious question is whether this spending makes sense. Microsoft’s relationship with OpenAI gives you a window into how these investments flow through to earnings and what the financial logic actually looks like when you pull it apart. The short answer is that the revenue side is growing fast but still well behind the infrastructure spend. Dan Ives at Wedbush described 2026 as “the year AI spending must start showing returns.” The market’s patience is real but not unlimited.

A year ago, the AI model race felt like a US competition with China watching from the sidelines. That framing is now outdated.

In May 2026, four Chinese labs released competitive open-weight coding models within a 12-day window. Z.ai’s GLM-5.1, MiniMax M2.7, Moonshot’s Kimi K2.6, and DeepSeek V4 all landed at roughly the same capability ceiling, and none costs more than a third of comparable Western frontier models per inference. These are not cheap imitations. They are genuinely competitive products at dramatically lower cost.

On the Western side, April 2026 saw GPT-5.5, Claude Opus 4.7, and DeepSeek V4 all launch within days of each other. The competitive cycle, which once moved in annual increments, has compressed to weeks. Every time a lab releases a model, another lab is days away from responding.

OpenAI is on track to hit $20 billion in annualised revenue in 2026, up from $3.7 billion the year before. Anthropic started 2025 at a $1 billion run rate and hit $7 billion by October. Both are preparing for IPOs. The structural shift here is that these companies have stopped being research labs that also do business and started being businesses that also do research. That changes how they prioritise, what they build, and who they compete with.

One of the most interesting subplots is the browser. The browser war has returned, and this time AI is the weapon of choice. The interface layer is becoming a significant battleground because whoever controls how people access AI controls the monetisation. This is not a coincidence that Microsoft, Google, and several AI-native startups are all investing heavily in browser-level AI simultaneously.

The Stanford 2026 AI Index is the most comprehensive data source we have on where AI actually stands in the real world, and it tells a nuanced story.

The headline number is impressive: generative AI is now in use at 70% of organisations globally in at least one business function. The productivity data is real too. Studies show 14-15% gains in customer support, 26% in software development, and 50% in marketing output. These are not trivial numbers.

But the production reality is considerably more modest than the headline suggests. Only 11% of organisations are actively running agentic AI systems in production. Another 14% have solutions in development. The rest are still in the exploration or pilot phase. The gap between “using AI somewhere” and “running AI as core infrastructure” is enormous, and most companies are still in that gap.

This maps directly to what has happened to the software industry. The SaaS model is under genuine pressure from AI, and the companies that understand this are moving fast. The ones that don’t are discovering that their competitive moat is thinner than they thought. Gartner projects that 40% of enterprise applications will embed task-specific AI agents by the end of 2026, up from less than 5% a year earlier. That is a massive structural shift in how enterprise software works.

This is the part of the AI story that does not get enough attention in mainstream coverage, so let’s be direct about it.

Training a single frontier model in 2026 generates an estimated 72,000 tonnes of CO2-equivalent emissions. That is up from 5,184 tonnes for GPT-4 and 8,930 tonnes for Meta’s Llama 3.1. The emissions trajectory of AI model training has roughly doubled every major model generation. Global data centre electricity consumption is on track to double between 2022 and 2026, according to the IEA.

Every major AI company has signed nuclear power deals. Microsoft restarted Three Mile Island specifically to power its AI data centres. Amazon, Google, and Meta have all made similar commitments. This is not greenwashing. It is a genuine acknowledgement that the only energy source capable of delivering reliable, around-the-clock power at the scale AI requires is nuclear.

The hardware layer is equally constrained. High Bandwidth Memory, required for all frontier AI GPUs, remains in tight supply. The combination of power constraints, memory constraints, and data centre construction timelines means that even $650 billion in committed spending cannot be deployed instantaneously. The infrastructure buildout will stretch across multiple quarters as supply catches up to demand.

Local governments in the United States are beginning to push back. Some have embraced restrictions or outright bans on new data centre development, citing power grid pressure and water usage for cooling. This is a new and underappreciated risk to the AI buildout timeline.

Nvidia remains the dominant force in AI hardware. Its data centre revenue continues to grow faster than custom chip alternatives, which means hyperscaler demand for third-party GPUs is increasing even as those companies invest in their own silicon. Amazon’s Trainium chips, Google’s TPUs, Microsoft’s Maia, and Meta’s MTIA are all real products used in production. But none has displaced Nvidia at the frontier. They complement it, they reduce dependency at the margin, but the fundamental Nvidia position has not changed.

What has changed is the supply chain. Companies like Super Micro Computer sit at the centre of the AI hardware buildout, assembling the server systems that house those GPUs and ship them to data centres at scale. The infrastructure layer is where a lot of the less-discussed value creation in the AI economy is happening. When people talk about AI, they talk about models and applications. The companies actually building the physical infrastructure are quieter but equally important to the story.

Three things matter most over the next 90 days.

Watch #1 — The IPO Filings

OpenAI has been laying groundwork for an IPO that could value it at $1 trillion. Anthropic has hired Wilson Sonsini to prepare for a potential listing. If either files in Q3, it will be one of the most significant market events of the year. The valuations being discussed would make these among the largest IPOs in technology history, and the filings will give the public the first real look at the economics of frontier AI at scale.

Watch #2 — The Returns Narrative

Several major AI infrastructure projects are reaching the point where the companies that funded them need to start explaining returns to shareholders. Microsoft, Google, and Amazon will all report Q2 earnings in July, and analysts will push hard on the question of whether AI revenue is growing fast enough to justify the capex. If the answers are unconvincing, sentiment could shift quickly.

Watch #3 — The Chinese Model Acceleration

Four competitive models in 12 days was a statement. The question is whether Western labs respond by accelerating their own release cadences, and whether the cost gap between Chinese and Western models starts influencing enterprise procurement decisions at scale. If enterprises start choosing cheaper Chinese models for non-sensitive workloads, that changes the economics of the whole industry.

Data in this report draws from the Stanford 2026 AI Index, Goldman Sachs AI capex modelling, company earnings reports, and public filings. All figures accurate as of June 2026.

This report represents the editorial opinion of AllinAllSpace and does not constitute financial or investment advice. AllinAllSpace is not a registered investment advisor.