Gartner: $4 trillion in AI spending in 2027, including $2.4 trillion on infrastructure alone

Gartner: $4 trillion in AI spending in 2027, including $2.4 trillion on infrastructure alone

🔎 $4,050 billion: AI enters its industrial phase

On November 12, 2026, Gartner published a forecast that deserves more than a quick press-release skim: $4,050 billion in AI spending worldwide in 2027, up 47.5% from $2,746 billion in 2026. More than $1,300 billion added in twelve months, according to Gartner.

But the figure that really matters is not the total. It's the breakdown. 59% of that sum — $2,397 billion, up 53% — will go into infrastructure: data centers, servers, networking, energy. By contrast, generative AI models account for only $51 billion. Forty-seven times less.

Why now? Because Gartner added $413 billion to its 2027 forecast in a single quarter, because it is already projecting a slowdown to +28.3% in 2028, and because, in the same week, it reminded everyone that more than $3,500 billion in investments could end up written off by 2030. The real economy of AI looks less and less like the launch demos. This is no longer a race for models: it's heavy industry.

The essentials

• $4,050B in global AI spending in 2027 (+47.5% vs $2,746B in 2026), according to Gartner (November 2026).
• AI infrastructure captures $2,397B, or 59% of the total (+53%). Optimized AI servers, the market's largest segment, reach $957.2B (+81%).
• GenAI models account for just $51B (+82%), and foundation models retain 81% of that market.
• AI agents and assistants double: $30B in 2026, $67B in 2027 — the fastest-growing category.
• Gartner added $413B to its forecast in a single quarter, almost all of it toward infrastructure, and already anticipates a slowdown to +28.3% in 2028.
• The flip side: more than $3,500B in investments at risk of being written off by 2030, and over 80% of neoclouds that could lose their independence.

Recommended tools

You don't have $2,400B in capex? Rest assured: that's not the right strategy. The right one is to rent the infrastructure that others are building.

The rented AI compute market — AI-optimized IaaS — will grow from $66B in 2027 to $174B in 2030 according to Gartner (ICTjournal, November 2026). It's the realistic entry point for everyone except the hyperscalers.

$4,050B in 2027: the detail that matters more than the total

Global AI spending will reach $4,050B in 2027, but it's the breakdown of that amount that tells the story: infrastructure crushes everything else, software is progressing, and models remain a niche market.

First reading: AI-optimized servers become the market's largest segment, at $957.2B. Gartner attributes this surge to hyperscalers, neoclouds, and "memflation" — the rise in server prices driven by memory costs. People are buying more machines, and each machine costs more.

Second reading: software ($649B) and services ($746B) are growing fast, but markedly slower than hardware. In 2027, AI remains a heavy industry before it is a software industry. Even AI cybersecurity ($86B, +69%) is growing more slowly than servers.

Third reading, the most revealing: of the $413B added to the 2027 forecast in a single quarter, $420B went to infrastructure — and $8B was taken away from software (Software Strategies Blog, November 2, 2026). When forecasters revise, they revise toward concrete and silicon, not toward code.

Speaking to Reuters, via Onvista, John-David Lovelock, chief of research at Gartner, summed up the mood: "user demand is evolving faster than expected." Translation: even the forecasters are running behind the curve.

Why infrastructure is swallowing 59% of the pie

Because every AI dollar rests on hardware: both training and inference run on servers, in data centers, powered by electricity — and each of these line items is becoming an investment sinkhole.

First driver: "memflation." This is Gartner's term for the rise in server prices tied to memory — the DRAM and HBM that power AI systems. The result: the optimized AI server segment jumps from $527.6 billion to $957.2 billion in a single year. A mix of rising volumes and per-component inflation, which benefits manufacturers and penalizes all their customers.

Second driver: the expansion of workloads. Lovelock stresses this in the Reuters interview: inference and training workloads are spreading to entire categories of applications. Every query sent to a model costs compute — and therefore capital. AI is no longer a lab project; it's a production workload.

Third driver: the capacity race is widening. The hyperscalers (AWS, Google, Microsoft) are no longer the only ones stacking up GPUs. ByteDance is reportedly preparing $70 billion in AI capex for 2026: China is stepping squarely into the infrastructure battle, as we detail in our analysis of Chinese capex. On the American side, SpaceX and Reflection AI have sealed a $6.3 billion deal that makes Colossus 2 the "AWS of AI", with Musk locking down frontier infrastructure. When a rocket maker and a social network become compute providers, it means demand is durably outpacing supply.

And opposition to data centers, so visible in the United States? A secondary factor, according to Lovelock: "when one door closes, another opens." Projects move; they don't disappear.

The real winners aren't the model labs

The big beneficiaries of the $4,050B in 2027 aren't the ones building the models, but the ones selling the building blocks: servers, memory, energy, data center real estate, network equipment.

The ratio is brutal: $2,397B in infrastructure versus $51B in GenAI models — a ratio of 1 to 47. For every dollar spent on models, forty-seven go into the pipes. Faced with that math, the "model as the product" narrative is hard to sustain.

Even within the model market itself, concentration is severe. Domain-specific models show the fastest growth (+99%), but foundation models still hold 81% of the market. Concretely: OpenAI's GPT-5.5, Google's Gemini 3.1 Pro, or Anthropic's Claude Opus 4.7 capture the bulk of the budgets, while vertical models fight over what's left.

My take? This is AI's "AWS moment," but one floor down. In the 2000s, the most durable winners of e-commerce weren't always the stores: they were often the sellers of servers and logistics. In 2027, the durable winners of AI are the sellers of memory, electricity, and white space. The model labs, for their part, are absorbing costs they don't control — memflation hits their margins before it hits their suppliers'.

AI agents and assistants: the category that doubles every year

At $67B in 2027 versus $30B in 2026, AI agents and assistants form the fastest-growing category in Gartner's entire forecast — and the only software layer that nearly doubles year over year.

The explanation fits in one sentence: companies are moving from copilots, which suggest, to agents, which execute. Cutting-edge models are built for this shift — GPT-5.5, Claude Opus 4.7, or Gemini 3 Pro Deep Think are now evaluated first and foremost on agentic tasks — and enterprise demand is following: automating entire processes rather than drafting text.

Let's keep a cool head, though. $67B is spectacular growth but a modest share: 1.7% of the total. Agents are the visible face of spending whose bulk remains buried in infrastructure. It is nonetheless the layer where an SME can still capture value without billions in capital: agents are bought on subscription, not as data centers.

This is also the most actionable signal in Gartner's entire press release: the segment that is doubling is the one where usage is actually being deployed. If you're going to follow only one curve in this forecast, follow this one.

2028, the wall: +28.3% and write-downs in limbo

Gartner itself is predicting the slowdown: AI spending growth will drop from 47.5% in 2027 to around 28.3% in 2028. A sharp deceleration — even though it still represents more than $1,100B added in a single year.

In parallel, the October 2026 warning remains in effect and was confirmed in mid-November (Yahoo Finance): if the current investment pace holds, more than $3,500B of AI infrastructure could be written off by 2030. We break down this scenario and its mechanisms in this analysis.

The accounting relayed by ICTjournal (November 6, 2026) is even starker. To be profitable with a 12% return on capital, AI data centers would need to generate $8,500B in cumulative revenue by 2030. Yet attributable application revenues would not exceed $3,600B. Hence potential write-downs exceeding $4,000B by 2032.

The designated victims fit a profile: the neoclouds. More than 80% would cease operating independently by 2030, squeezed by costs and lacking differentiated business models. Lovelock remains measured: bankruptcy or acquisition. The hyperscalers, for their part, hold up thanks to their ancillary revenues — traditional cloud, software, advertising — which absorb AI capex.

An important nuance: these write-downs will do little to slow investment, again according to Lovelock. Demand is advancing faster than the losses, and local opposition to data centers remains a secondary factor. The 2028 wall will slow growth, not the race.

Finally, market context: these forecasts land in a stock market sector already shaken by the turbulence surrounding the Jensen Huang scandal (Yahoo Finance, November 2026). Psychology follows the numbers, in both directions.

Without $2.4T in capex: how to position yourself in practice

You won't be building a data center, and that's a good thing. The 2027 strategy comes down to three rules: rent the compute, bet on the layer that's doubling, and avoid fragile vendors.

Rule 1 — rent, don't build. AI-optimized IaaS, AI compute billed on usage, will grow from $66B in 2027 to $174B in 2030 (ICTjournal, November 2026). AWS, Azure, and Google Cloud cover most needs; to host the application around the model, Hostinger does the job starting at around €3 per month (November 2026, check hostinger.com). The giants carry the capex, everyone else pays opex: that's exactly the asymmetry you should exploit.

Rule 2 — bet where things double. AI agents and assistants (+123% in 2027) are the segment where a small organization can still carve out a position: integration, business verticals, process automation. The model market is locked up at 81% by the foundation model players; the market for agentic use cases isn't, yet.

Rule 3 — avoid dependence on the neoclouds. When more than 80% of a category risks losing its independence by 2030, building your stack on it is a dangerous bet. Multi-cloud, short contracts, data portability: three precautions that cost nothing today and everything tomorrow.

One last point of caution: memflation doesn't just hit the hyperscalers. Rising memory prices ripple up the entire chain, all the way to GPU instances rented by the hour. Budget your compute costs assuming prices rise, not fall.

❌ Common Mistakes

Mistake 1: Reading $4,050B as revenue

That's not what AI earns — it's what AI costs. Attributable application revenue would cap out at $3,600B cumulative by 2030, far short of the $8,500B needed to make the data centers profitable (ICTjournal, November 2026). The solution: think in terms of costs and returns, never in terms of hype.

Mistake 2: Believing the model market is the AI market

$51B in models versus $2,397B in infrastructure: a ratio of 1 to 47. Building a business, investing, or hiring "in AI" without looking at where the money actually goes means being on the wrong floor. The solution: analyze the full value chain — memory, servers, energy, networking — before the application layer.

Mistake 3: Treating a neocloud like a hyperscaler

Neoclouds don't have the ancillary revenues that absorb capex. More than 80% could cease operating independently by 2030 (Gartner, November 2026). The solution: short contracts, data portability, and a multi-cloud architecture from day one.

Mistake 4: Projecting 2027's growth onto 2028

Gartner is already forecasting +28.3% for 2028, versus +47.5% for 2027. Building a capacity plan or a business plan on the current slope means running straight into a wall. The solution: flexible commitments and two-speed scenarios.

❓ Frequently Asked Questions

How much will the world spend on AI in 2027?

$4,050 billion, up 47.5% from the $2,746 billion spent in 2026, according to Gartner (November 2026). Infrastructure captures 59% of that, i.e., $2,397 billion, including $957.2 billion for AI-optimized servers alone — the market's largest segment.

What is "memflation"?

It's Gartner's term for the rise in server prices driven by memory costs (DRAM, HBM). It partly explains the 81% jump in the AI-optimized server segment in 2027: volumes are growing, but each machine also costs more to produce.

Do AI agents really account for $67 billion?

Yes, according to Gartner's November 2026 forecast: $67 billion in 2027 versus $30 billion in 2026, i.e., +123%. It's the fastest-growing category in the entire forecast, driven by the shift from copilots — which suggest — to agents, which execute complete business tasks.

Should we talk about an AI bubble?

Gartner instead points to the risk of massive depreciation: more than $3,500 billion in investments could be written off by 2030, and more than $4,000 billion by 2032, according to calculations relayed by ICTjournal. But the firm also forecasts +28.3% growth in 2028: a slowdown, not a collapse.

Who really wins in this economy?

The infrastructure providers: server manufacturers, memory vendors, energy suppliers, data center operators. Hyperscalers are holding up thanks to their diversified revenues, neoclouds risk bankruptcy or acquisition, and model labs have only a $51 billion market to share among themselves.

What can an SME do with a normal budget?

Rent its computing rather than build it, invest in the agentic layer, which doubles every year, and avoid dependence on a fragile vendor. To host the applications around the model, solutions like Hostinger start at around €3 per month (November 2026).

✅ Conclusion

In 2027, AI will be less a software revolution than a heavy industry worth $4,050 billion, nearly 60% of which will go into concrete, silicon, and electricity: follow the money, not the demos. To anticipate the next installment of this saga, carry on with our analysis of the $3,500 billion of AI infrastructure at risk of being written off by 2030.