The race in artificial intelligence is no longer just an elegant competition of models, code, and benchmarks. It increasingly resembles an industrial war over electricity, memory, land, water, fiber, transformers, and chips. The new power is not measured solely by parameters or users but by contracted gigawatts and secured computing capacity years into the future.
Anthropic’s case summarizes this shift particularly well. According to Reuters, Google plans to invest up to $40 billion in the company, while Anthropic has committed to spending $200 billion on Google Cloud and chips over five years, as reported by The Information and also covered by Reuters. The operation follows an obvious financial logic: Google injects capital into one of the strongest AI companies and, at the same time, turns part of that investment into future revenue from its cloud and TPUs.
An uncomfortable question arises: if the money flowing into AI startups eventually returns to the major cloud providers through computing contracts, are we witnessing genuine value creation or a closed circuit of financing, capacity, and committed revenue? The answer isn’t simple. There is real demand, enterprise clients, and rapidly growing products. But there is also a physical dependency that can no longer be hidden behind the word “cloud.”
The Circular Economy of the Gigawatt
Anthropic isn’t just signing server contracts; it’s buying survival time in a race where running out of compute means losing users, delaying models, and ceding ground. Its agreement with Google and Broadcom will give access to several gigawatts of next-generation TPU capacity starting in 2027, according to the company. Other reports suggest that the total could be around 5 GW, including 3.5 GW linked to Broadcom.
Amazon has adopted a similar logic. In April, Anthropic announced a deal with AWS to secure up to 5 GW of capacity for training and deploying Claude, using Trainium2 and Trainium3 chips as the core elements. The Associated Press also reported a commitment of over $100 billion with AWS over ten years.
Microsoft and NVIDIA are also in the game. Anthropic committed to purchasing $30 billion of Azure capacity and up to an additional gigawatt with systems like NVIDIA Grace Blackwell and Vera Rubin, according to Microsoft’s joint announcement. Additionally, Anthropic announced a $50 billion investment in own data centers in the U.S. with Fluidstack, with facilities in Texas and New York.
| Agreement or Investment | Capacity or Announced Amount | Main Takeaway |
|---|---|---|
| Google / Broadcom / Anthropic | Up to $200 billion and several GW of TPU, per reports | Google invests and secures future cloud revenue |
| Amazon / Anthropic | Up to 5 GW and over $100 billion in AWS over ten years | AWS enhances Trainium as an alternative to NVIDIA |
| Microsoft / NVIDIA / Anthropic | $30 billion in Azure and up to 1 additional GW | Azure gains significant AI capacity outside of OpenAI |
| Fluidstack / Anthropic | $50 billion in US data centers | Anthropic seeks more controlled capacity |
| SpaceX / Anthropic | Over 300 MW and more than 220,000 NVIDIA GPUs | Urgency of compute breaks down rival boundaries |
Money doesn’t disappear; it circulates. Investors fund AI companies, these companies commit that capital to cloud capacity, hyperscalers justify more data centers, and chip manufacturers receive increasingly large orders. The circle works as long as demand grows, models improve, and clients pay. If any part cools down, the system must demonstrate what portion was sustainable business and what was anticipation of an uncertain future.
Memphis Demonstrates That Ideology Matters Less Than Compute
The most striking episode is the SpaceX agreement. In May, Anthropic announced a partnership to use SpaceX’s compute capacity, with over 300 MW and more than 220,000 NVIDIA GPUs available soon, enabling increased limits for Claude Code and the API. Reuters later reported, based on SpaceX’s S-1 filing, that Anthropic would pay $1.25 billion per month until May 2029 for access to the Colossus and Colossus II centers in Tennessee, with a termination option with 90 days’ notice.
It’s hard to ignore the irony. Elon Musk has directly competed against Anthropic with xAI and Grok, yet the urgency for capacity has made SpaceX a supplier to its rival. In the AI economy, principles matter less than available megawatts. An already built cluster is worth more than a promise of a data center in two years.
The agreement also serves to differentiate reality from narrative. The tangible part is in Memphis: 300 MW, GPUs, contracts, and rising usage limits. The more speculative part involves orbital data centers. Reuters reported that SpaceX’s pre-IPO documentation indicated their orbital computing and off-Earth industrialization initiatives are in early stages, involve untested technologies, and might not achieve commercial viability.
Space is attractive as a narrative, but engineering remains stubborn. In orbit, there’s no convection like on Earth: dissipating heat at large scale requires enormous radiators. Latency imposes physical constraints. And every hardware failure is far more costly to repair than in a terrestrial data center. For now, the critical data resides in Tennessee, not in space.
The Power Grid Starts to Pay the Price
The International Energy Agency estimates that global electricity consumption for data centers could double to around 945 TWh by 2030, representing about 3% of worldwide electricity. Between 2024 and 2030, data center electricity demand might grow at approximately 15% annually—more than four times faster than other sectors.
Goldman Sachs Research projects an even steeper increase: global data center electricity demand could rise by 50% in 2027 and up to 165% by the end of the decade relative to 2023 levels. AI is a key but not the only driver.
The most apparent stress point is PJM, the grid covering much of the eastern US. Reuters reported capacity prices in PJM surged from $28.92 per MW-day in 2024-2025 to $269.92 in 2025-2026 and hit a cap of $329.17 in 2026-2027, driven mainly by demand from data centers in regions like northern Virginia.
The IEEFA also highlighted that the proliferation of data centers is increasing electricity costs for consumers in the PJM region, with capacity prices multiplying over ten times from 2024-2025 levels. The issue isn’t just about generating more electricity; building lines, substations, transformers, cooling systems, and firm capacity with industrial timelines is essential, not just model rollout schedules.
This reflects the Jevons paradox: each efficiency improvement lowers operational costs but can lead to increased overall usage. More efficient models enable more users, agents, API calls, and automation. The outcome isn’t necessarily lower total consumption—often it’s higher demand.
Costs Also Hit Consumers
The pressure extends beyond electricity bills. Memory is another front. AI data centers consume vast amounts of high-performance HBM and DRAM. This demand competes with manufacturing capacity for traditional server, laptop, PC, and consumer device memory.
Avnet reports that AI and data center demand are driving significant increases in memory and storage components, with some parts experiencing 30-60% monthly growth and quarterly increases over 60% in server DRAM, based on market data cited by the company.
Although the average consumer doesn’t buy HBM to train models, they may end up paying more for DDR5, SSDs, or laptops because factories prioritize higher-margin products. AI not only consumes electricity but also reshapes the semiconductor supply chain — millions of consumers bear part of the cost of a infrastructure they will never see.
Cybersecurity Moves Into the Physical World
The link between AI, energy, and cybersecurity is increasingly delicate. Gartner predicts that by 2028, a poorly configured AI system in cyber-physical infrastructure could shut down a critical national infrastructure in a G20 country. The firm advocates for safe override methods to maintain ultimate human control over critical systems.
It’s not just about hackers. Risks also stem from autonomous systems misconfigured, contaminated training data, faulty sensors, or agents executing dangerous recommendations on electrical networks, industrial plants, or transportation systems. When AI stops just answering questions and begins acting physically, the margin for error changes.
Anthropic provides a paradoxical example. Its Mythos Preview model, used within Project Glasswing, has revealed old and severe vulnerabilities, including a 27-year-old bug in OpenBSD and a 16-year-old one in FFmpeg, according to the company. Project Glasswing involves actors like AWS, Apple, Google, Microsoft, NVIDIA, CrowdStrike, and the Linux Foundation working on defensive applications.
The good news: AI can help identify vulnerabilities before attackers do. The bad news: the bottleneck shifts to patching, verification, and governance. Detecting vulnerabilities faster is useless if critical systems cannot be fixed at the same pace.
AI Is No Longer Just Software
For years, cloud computing was sold as an almost weightless entity. Data “went up,” applications “lived” above, and infrastructure was outside the story. AI has broken that illusion. Every token generated depends on physical chips, physical memory, physical electricity, water or air cooling, and electrical grids with very real limits.
The gigawatt war doesn’t mean AI will fail; it means its expansion will have more visible costs. It will require energy, long-term agreements, new data center designs, better chips, increased efficiency, greater transparency, and a less naive public discussion about who pays for infrastructure.
AI can boost productivity, science, software, healthcare, education, and manufacturing—but it doesn’t float in the air. It lives in windowless buildings connected to high-voltage lines, filled with chips converting electricity into calculation and heat. The next phase of AI development won’t only happen in labs; it will also happen at substations, memory factories, power markets, and control rooms.
The vapor rising from a data center isn’t just a metaphor; it’s the thermal footprint of artificial thought.
Frequently Asked Questions
Why is there talk of a “gigawatt war” in AI?
Because major AI companies now compete not only for models or talent but for securing several gigawatts of electrical and computational capacity to train and run their systems.
What makes the Anthropic case special?
Anthropic has closed large-scale agreements with Google, Amazon, Microsoft, NVIDIA, Fluidstack, and SpaceX. Its growth depends directly on securing compute capacity for years.
Can AI increase electricity and component costs?
Yes. The demand from data centers strains local power grids, and the need for memory for AI can displace manufacturing capacity from traditional consumer products.
Are space-based data centers viable?
Today, they are highly speculative. SpaceX has acknowledged in pre-IPO documentation that its orbital compute initiatives involve untested technologies and may not be commercially viable.
via: LinkedIn