Gartner: spend outruns revenue
A 12% return on planned AI infrastructure would need ~$8.5T of revenue to 2030; Gartner sees at most $3.6T. Hyperscalers can absorb it — smaller GPU clouds may not.
Chips, data centres, energy and cloud capacity: the physical layer that decides who can train and serve AI, and at what cost.
Free. Weekly. One-click unsubscribe. See our privacy policy.
Amazon (~$220B) and Alphabet ($195–205B) keep raising capex and power is the binding constraint. This week the question shifted to returns: Gartner sees nearly $1T of AI infrastructure spend in 2026 and over $3.5T at risk of write-downs by 2030, with neoclouds most exposed. Meanwhile the chip map is fragmenting — Huawei has 28% of China's AI accelerators against Nvidia's 34%, and Microsoft, ByteDance and others push custom silicon. Expect capacity to stay tight into 2027, and some weaker operators to stumble.
A 12% return on planned AI infrastructure would need ~$8.5T of revenue to 2030; Gartner sees at most $3.6T. Hyperscalers can absorb it — smaller GPU clouds may not.
Microsoft's Maia 300 promises 40% more tokens per megawatt than GPUs from 2027, and ByteDance is negotiating custom chips with MediaTek — inference is where buyers diversify first.
McKinsey's estimate turns energy into the key variable of AI cost; US data-centre land now prices on power access (~$6B bought in H1).
The first US small modular reactor permit (TVA, 300 MW) and LS Power's record $6B fund show capital moving from chips to electrons.
DeepSeek and Huawei open-sourced Ascend programming tools to reduce CUDA dependence — a parallel AI stack is forming.
Lock in AI capacity and prices for 12 months where you can, keep a second provider ready, and ask for the data-centre region in every AI contract. If you buy from a smaller GPU cloud, check its funding and have an exit plan.
Trackers stay on screen until a newer, verified value replaces them.
Source: Hargreaves Lansdown · Data Center Knowledge · Gartner · as of Oct 16, 2026
Company guidance for calendar 2026 (midpoint for ranges). Microsoft reports on a June fiscal year and is not comparable.
Source: Fortune (Amazon) · Nasdaq (Alphabet) · ValueAdd VC (Meta) · as of Oct 11, 2026
Signals scored 9+ or structural shifts stay here for 90 days.
Data-centre revenue rose 117% year on year and management called next year's outlook supply-constrained. AI infrastructure demand shows no sign of slowing, which keeps GPU prices — and AI costs — firm.
Higher memory costs pushed the figure up from about $200B; AWS revenue grew 37% to $42.2B. Cloud AI capacity stays tight into 2027.
The range was $180–190B before. Together with Amazon and Meta, Big Tech's AI build-out is running well above $500B this year.
Badges show how long each signal stays: this week, this month, key context or archive.
IDC data for the first half of 2026: US export controls limited Nvidia's supply while domestic accelerators gained ground. China's parallel AI stack is now measurable in market share.
AI providers spend nearly $1T in 2026 and another $6T to 2030; a 12% return would need about $8.5T of revenue, but Gartner sees at most $3.6T. Neoclouds are the most exposed; hyperscalers can absorb it.
The British AI-cloud company's Norwegian site slips as it prepares to list — execution risk at neoclouds is real.
The Series B funds optical circuit switching for AI clusters — networking is the next constraint after chips and power.
Faster than the 108% growth of the first nine months, a sign of China's growing weight in the semiconductor supply chain.
Production deployment is planned for 2027; Maia 200 already powers Copilot and Teams in US, European and Australian data centres. Hyperscalers are moving inference onto their own silicon.
Advanced talks would make the TikTok owner a major ASIC customer — another large AI buyer diversifying away from Nvidia.
Its CEO says demand could exceed capacity despite large expansion plans — the lithography bottleneck caps how fast AI chips can scale.
AI demand keeps lifting the foundry; it also confirmed talks about a role in Elon Musk's Texas 'Terafab', insisting on operational control.
Access to power now sets land value more than location. Energy, not plots, is what developers are really buying.
Operators are turning to on-site generation, while uncertainty beyond 2030 makes it hard to size new builds. Expect energy to show up in AI pricing.
The release includes compute and chip-to-chip communication libraries and Ascend support for TileLang. China's AI stack is closing ranks around domestic chips, which matters for anyone pricing Chinese open models.
Nuclear SMRs move from slides to permits — a long-term answer to AI's power hunger, with GE Vernova Hitachi's boiling-water design.
Its first debt financing pays for hardware at Bruyères-le-Châtel, run by Eclairion. Mistral aims for 200 MW of capacity across Europe by the end of 2027.
Spain is preparing its bid for the European Commission's AI gigafactory call, with candidate sites such as Móra la Nova and San Fernando de Henares. Winning one would put frontier-scale compute in Spain.
The range was $180–190B before. Together with Amazon and Meta, Big Tech's AI build-out is running well above $500B this year.
Higher memory costs pushed the figure up from about $200B; AWS revenue grew 37% to $42.2B. Cloud AI capacity stays tight into 2027.
Hosted by NAISS with RISE, it is one of 31 EuroHPC AI Factory sites. Startups and SMEs can apply for subsidised compute — worth knowing if you build models in Europe.
OpenAI and Sur Energy announced up to $25B and 500 MW in October 2025; the project is still being defined and later reports cite lower figures. Treat headline AI-campus numbers as options, not capacity.
Hyperscalers are co-locating with large power sources, turning former power-plant sites into AI campuses.
A second 500 MW phase arrives a year early: former crypto miners keep pivoting into AI hosting.
Labs are building their own infrastructure teams — and Europe is on Anthropic's capacity map.
Power, not chips, is becoming the bottleneck. Infrastructure investors are positioning upstream of AI capacity.
Education, job training, water and energy conservation: social licence is now part of the cost of AI capacity.
Compute and energy prices flow straight into the price you pay for AI. Capacity announcements tell you where costs, latency and data-residency options are heading.
Tell us what slows your business down. In 30 minutes we'll show you where AI pays back first — and what it costs.