$650–700B for AI Compute: The Capex Race and Enterprise

2026 is not “another year of AI”. It is the year when AI finally meets physics: electricity, cooling, optics, supply chain, and data centre real estate. According to financial projections and company commentary, four hyperscalers (Amazon, Alphabet/Google, Microsoft, and Meta) plan to spend a combined approximately $650-700 billion in capital expenditures (capex) — largely on AI compute, data centres, and networking. For enterprise, this means one thing: compute is becoming a strategic commodity and your architecture and procurement must change.

The Capex Race in Numbers: Why We Are Suddenly Talking About Hundreds of Billions¶

In recent weeks, two similar figures keep appearing: $650B (a more conservative sum) and nearly $700B (an aggregation according to some analyses). This is not marketing — these are specific capex outlooks and their extrapolations.

• Amazon announced a projection of $200B capex for 2026 across “AI, chips, robotics, and low earth orbit satellites”. (TechCrunch, 5 Feb 2026)
• Alphabet (Google) is talking about $175-185B capex for 2026, significantly above 2025. (TechCrunch, 5 Feb 2026)
• Meta expects $115-135B capex for 2026. (CNBC, 6 Feb 2026; TechCrunch, 5 Feb 2026)
• Microsoft did not have a unified annual outlook in the cited materials, but quarterly capex of around $37.5B implies roughly ~$150B/year at a similar pace. (TechCrunch, 5 Feb 2026)

Why should you care? Because hyperscalers at this pace are changing cloud pricing, GPU capacity availability, region priorities, and ultimately who will be able to run AI at scale.

Compute Is Not Just “GPU”. It Is a Data Centre Stack — and It Has Bottlenecks¶

When people say “AI infrastructure”, most imagine GPUs. The reality in 2026 is more like a data centre supply chain: power connections, substations, transformers, distribution boards, UPS, cooling (liquid), optical transceivers, network fabric, and physical construction. This is precisely why capex is growing faster than spending on the chips themselves.

From an enterprise perspective, this has two consequences:

1. Regional capacity inequality will increase. Not all regions have the same power connection and construction capabilities.
2. “On-demand” will no longer be the default. For certain GPU/accelerator types, a “capacity reservation” mentality is practically returning.

What Investors Are Saying: Capex Growth = Pressure on Free Cash Flow¶

It is not just about who builds more data centres. Wall Street is watching what this does to cash. CNBC reports that the four largest internet companies in the US together generated approximately $200B free cash flow in 2025, which is a decline from $237B in 2024 — and a more dramatic drop may yet come due to AI investments. (CNBC, 6 Feb 2026)

In the same article, CNBC cites specific estimates:

• For Amazon, Morgan Stanley according to CNBC expects negative free cash flow of approximately -$17B in 2026, while Bank of America sees a deficit of up to -$28B.
• For Alphabet, Pivotal Research according to CNBC projects a free cash flow decline of nearly 90% to $8.2B from $73.3B in 2025.

Silicon Republic also points to investor nervousness and notes that the predicted capex would represent approximately 60% growth compared to $410B in 2025. (Silicon Republic, 6 Feb 2026)

What Impact This Will Have on Enterprise: 5 Practical Scenarios¶

1) Cloud Pricing: The “AI Tax” Will Not Be Called AI Tax¶

Hyperscalers need returns. But “we are raising GPU prices” sounds bad. In practice, expect a combination of: discount changes, pressure on commitment (Savings Plans/Reserved capacity), bundling (model + runtime + observability), and a larger price differential between regions.

2) Lead Times and Capacity Reservations: Procurement Returns as an Engineering Problem¶

If you are planning the rollout of a large inference platform (e.g., customer support, internal copilots, document workflows), you will be dealing with capacity before the model. For some workloads, it makes sense to negotiate in advance:

• commitment on GPU nodes / reserved capacity,
• multi-cloud fallback (at least 2 providers),
• alternative instance types and “acceptable performance tiers”.

3) Model Selection Will Be More About TCO Than Benchmarks¶

For enterprise use cases, it often does not matter whether a model wins by 2 points on a benchmark. What matters is whether it can be reliably operated with guaranteed latency and predictable pricing. This leads to greater adoption of patterns like model routing (a cheaper model as the default, a more expensive one only for hard queries).

`# Model routing (pseudo) — cost-first fallback

$650–700 Billion for AI Compute: What the Capex Race Means for Enterprise in 2026¶

2) escalate to strong model only when needed¶

def route(prompt, risk, sla_ms): if risk == “high” or sla_ms < 800: return “strong-model” if len(prompt) > 6000: return “strong-model” return “cheap-model”

Guardrail: budget per user/session¶

MAX_USD_PER_SESSION = 0.50`

4) Inference Architecture: Batching, Quantization, Caching as Standard¶

“We will pay for it” stops working when adoption scales. In 2026, we see enterprise teams standardizing:

• prompt caching (stable system prompts + parts of conversations),
• response caching for FAQ and repeated queries,
• batching for non-interactive workloads (summarization, classification),
• quantization and smaller SLMs for “edge” or private workloads.

If you want to go deeper on costs, we also recommend our internal perspective on AI cost optimization.

5) Security and Regulation: More Data, More Logs, More Audits¶

More AI traffic = more sensitive data in prompts, more integrations, more identity touchpoints. If you are building agentic workflows, this connects directly to topics from the article AI Agent Security in Enterprise.

Secondary Effects: Pressure on Productivity and Changes in Hiring¶

The capex race also has social impacts. CBS News reports that in 2025, companies directly referenced AI when announcing 55,000 layoffs — more than 12x compared to two years earlier (data from Challenger, Gray & Christmas). Some economists note that companies may sometimes use AI as a “narrative” for layoffs and overhiring corrections. (CBS News, 3 Feb 2026)

For CTOs, one practical takeaway emerges: AI is becoming part of the company’s operating model — not just an IT project. And the operating model changes through processes, tooling, security, and impact measurement.

What to Do in Practice: 90-Day Plan for CTOs¶

Step 1 — Establish an AI Capacity & Cost Baseline (weeks 1-2)¶

• How many requests / tokens / GPU hours do you actually consume today?
• Which workloads are interactive vs. batch?
• What is “mission-critical” (SLA) and what can degrade?

Step 2 — Introduce AI FinOps (weeks 2-6)¶

• budget per team / per product,
• cost anomaly detection,
• standard patterns: caching, routing, batching.

Step 3 — Negotiate Capacity and Plan for Vendor Risk (weeks 4-10)¶

In an environment where hyperscalers themselves are dealing with capacity pressures, it is sensible to have:

• at least 2 runtime paths (e.g., managed API + own inference),
• an exit plan for critical use cases (model + data + orchestration),
• options for sovereign cloud and regulated data domains.

Step 4 — Build a Platform, Not Just a POC (weeks 6-12)¶

If you want “AI in the company”, you need platform engineering: identity, secrets, observability, deployment. It is no coincidence that the topic of Internal Developer Platform is returning in enterprise. At CORE SYSTEMS, we typically integrate this into the work of the Cloud & Platform and AI & Agentic Systems teams.

Conclusion: AI Strategy Is Now Infrastructure + Finance + Architecture¶

$650-700B capex is not just an interesting fact from earnings calls. It is a signal that AI is moving into a phase where those who can build and operate reliable infrastructure while keeping TCO under control will win.

If you want to scale AI, address: capacity planning, AI FinOps, security, and platform engineering. We are happy to walk you through the architecture, vendor strategy, and a concrete roadmap.

Sources: TechCrunch (5 Feb 2026), CNBC (6 Feb 2026), Silicon Republic (6 Feb 2026), CBS News (3 Feb 2026). Cited figures are based on the referenced articles and their linked earnings releases.