Three things this audience took for granted six months ago:
Subsea cables transit the world’s chokepoints for free
Anthropic and xAI are competitors, not partners
Data centres stay on no matter what
This week, Iran demanded rent on the cables. The Virginia grid operator won the right to force data centres offline. And SpaceX’s IPO filing revealed how much Anthropic is paying to rent the GPU clusters Musk built for xAI to beat them.
Assumptions expire fast in this market.
I’m Ben Baldieri, and every week I break down the moves shaping GPU compute, AI infrastructure, and the data centres that power it all.
Here’s what’s inside this week:
Let’s get into it.
Data transfer via subsea cables may soon not be free.
Iran announced it will impose fees on undersea cables running through the Strait of Hormuz, naming Meta, Google, Amazon, and Microsoft. State media claimed exclusive rights to repair and maintain the cables. Repair ships have been unable to enter since the war began in February. The four hyperscalers are now buying into an overland fibre route through Iraq built by IQ Networks alongside pipeline corridors.
Why this matters:
The governing treaty is the 1884 Convention for the Protection of Submarine Telegraph Cables, designed for telegraph wires and barely updated since. The FALCON and Gulf Bridge cables transit Iranian territorial waters. Iran is exploiting a grey zone that has existed for 140 years because nobody expected the cables to matter this much.
Prior to 2012, hyperscalers used less than 10% of international subsea capacity. By 2024, Google, Meta, Microsoft, and Amazon controlled 71%. Google alone has invested in 33 cables. These are private lines with no common carriage obligations. Four American companies own the majority of the physical internet, and a country at war is claiming the right to tax the portion running through its waters.
If Iran succeeds in imposing fees or licensing, it sets a precedent that any coastal state can charge rent on the cables transiting its waters. Indonesia, Egypt, Singapore, and dozens of other chokepoint nations would take notice. The 1884 treaty assumed cables were public goods that benefited everyone. The hyperscalers turned them into private infrastructure that benefits four companies.
SpaceX’s IPO Filing Reveals Anthropic Is Paying $1.25 Billion Per Month for xAI Data Centre Space
xAI has gone from one of the leading AI labs to one of the largest neoclouds.
SpaceX’s IPO filing disclosed that Anthropic will pay SpaceX $1.25 billion per month for data centre space at xAI facilities, with the contract running to May 2029 and a 90-day termination clause. The same filing shows xAI lost $2.4 billion in Q1 2026, driven by $7.7 billion in AI infrastructure spending in a single quarter. Colossus claims 1GW of nameplate capacity but operates at 330MW.
Why this matters:
We covered this deal in Issue #105. Now we have the price: $1.25 billion per month, $15 billion annualised. Anthropic’s revenue hit $30 billion in March (Issue #103). Half is committed to a single lease.
xAI spent $7.7 billion in one quarter and lost $2.4 billion. Grok trails every frontier coding model. Musk built the largest GPU cluster on earth to win the AI race. The business model that emerged is leasing it to the company that did.
1GW nameplate versus 330MW operational. The filing confirms what the market suspected: Colossus was announced at a scale it has not yet reached.
Blackstone and Google Launch the First Neocloud That Doesn’t Run on GPUs
When is a neocloud not a neocloud? When it is run by Google silicon, Blackstone capital, and Meta is the first customer.
Blackstone and Google announced a joint venture to sell Google TPU capacity as a standalone cloud. Blackstone committed $5 billion in equity. Google supplies TPUs, JAX/XLA, and managed services. Meta is the first customer. Initial target: 500MW. The new U.S.-based company will sell Google TPUs as compute-as-a-service distinct from Google Cloud’s own TPU access. Benjamin Treynor Sloss, a 20-year Google infrastructure executive, has been named CEO.
Why this matters:
Anthropic expanded its TPU capacity to 1GW in October 2025, then to 5GW through Broadcom in April 2026. Claude Opus 4.5 launched a month after the first expansion and broke every coding benchmark. Correlation is not causation, but the timeline is worth noting.
Google TPUs were previously available only through Google Cloud. Enterprise and sovereign buyers can now procure Google silicon without buying Google’s cloud. That opens a procurement channel that makes TPUs a direct competitor to NVIDIA GPUs in a way they were not when they were locked inside one company’s cloud.
Cerebras Runs a Trillion-Parameter Model at ~1,000 Tokens Per Second
Last week the company raised $5.5 billion and nearly doubled on day one, and this week it proved the silicon works.
Cerebras announced it served Moonshot AI’s Kimi K2.6, a trillion-parameter model, at approximately 1,000 tokens per second on its wafer-scale hardware, nearly 7x faster than GPU-based cloud alternatives. Artificial Analysis measured the system at 981 tokens per second. 6.7x faster than the next-fastest GPU-based cloud and 23x faster than the median inference provider. End-to-end response time on a 10,000-token prompt dropped from 163.7 seconds on the official Kimi endpoint to 5.6 seconds on Cerebras.
Why this matters:
Running the same model faster does not make it smarter, but it makes previously impractical workloads possible.
At 40 tok/s on GPUs, a 100-million-token output takes a month. At 1,000 tok/s, under two days. That gap opens an entirely new class of use cases that GPUs simply cannot serve.
The caveat: NVIDIA has a $6 trillion market cap, CUDA, and every hyperscaler as a customer. Faster tokens alone do not displace that. Groq tried and got acquired for $20 billion (Issue #80). The history of chip companies that were faster on paper is long.
AMD Commits $10B to Taiwan to Scale Helios Supply Chain
AMD is bringing the fight to NVIDIA.
AMD announced more than $10 billion in Taiwan investments to scale advanced packaging for its Helios rack-scale platform. The deal was signed with ODMs Sanmina, Wiwynn, Wistron, and Inventec. If Helios ships on time in 2H 2026, AMD enters the same delivery window as NVIDIA’s Vera Rubin.
Why this matters:
AMD’s bottleneck has never been silicon design. It has been packaging at hyperscaler volumes. $10 billion in Taiwan with four ODM partners targets that directly.
AMD held single-digit data centre GPU share eighteen months ago. Meta signed a 6GW deal (Issue #93). HUMAIN last year adds 500MW. If Helios ships on time, AMD could cross double-digit share for the first time.
Still a fraction of NVIDIA, but enough to give procurement teams a second option.
Google Pledges $15 Billion for a 1.2GW Campus in New Florence, Missouri
Missouri is quietly becoming an AI compute corridor.
Google announced a $15 billion investment in a 1.2GW data centre campus in New Florence, Missouri, with grid expansion commitments alongside utility Ameren Missouri. Nebius also broke ground on a separate 1.2GW campus in Independence, Missouri last week (Issue #106 EE).
Why this matters:
Two 1.2GW campuses from two different companies in the same state within two weeks. Missouri has cheap power, available land, and utilities willing to expand the grid. A year ago it was not on any data centre market map.
Ameren Missouri is expanding grid infrastructure to accommodate the load. PJM, which covers Northern Virginia, is doing the opposite (see below).
Google’s $15 billion sits directly adjacent to Amazon’s previously approved $35 billion campus on the same Montgomery County corridor. Combined with Nebius’s 1.2GW Independence site from last week, Missouri now hosts three of the largest single-project compute commitments announced in North America this quarter.
PJM Wins Emergency Approval to Force Data Centres Offline During Heat Events
The grid operator that serves Virginia, the largest data centre market on earth, can now shut you down after you have built.
PJM Interconnection received emergency approval to curtail data centre loads during extreme heat events. PJM manages the grid across 13 US states and the District of Columbia, covering Virginia, Pennsylvania, Ohio, and other major data centre markets. The RTO petitioned the DOE on Sunday May 17, citing planned plant outages and forecast peak demand that would leave less than 5.8GW of reserves. Curtailment is treated as a last resort before rolling blackouts and applies only to large loads with on-site backup generation, including hyperscale data centres.
Why this matters:
The permitting fight has been about whether you can build. This is about whether the grid can take your power away after you have built. Every operator in PJM territory, which includes Northern Virginia (the largest data centre market on the planet), now faces a risk that did not exist last month: involuntary curtailment during peak demand.
Data centre SLAs guarantee uptime. If the grid can force you offline, you either need backup generation to maintain the SLA or you renegotiate downward. Insurers pricing business interruption coverage in PJM territory now have a new variable. Financial services, healthcare, and government customers will ask whether Virginia is still the right market for workloads that cannot go down.
Google and Nebius are building 2.4GW in Missouri because the grid there is expanding to meet them. PJM is contracting.
The Rundown
Three assumed truths were proven false this week.
Subsea cables transit chokepoints for free. They don't anymore. The 142-year-old treaty governing them explicitly exempts itself during armed conflict — and there is currently armed conflict in the Strait of Hormuz. Iran is testing what happens next.
Anthropic and xAI are competitors. They aren't. Anthropic is now the largest customer for the cluster Musk built specifically to beat Anthropic. The model race collapsed into a real estate relationship.
Data centres stay on no matter what. They won't anymore. PJM acquired the legal authority over the weekend, and the contagion runs through insurance, site-selection economics, and the SLA framework every hyperscaler customer contract references.
The same pattern runs through the other four stories. Blackstone built a TPU procurement channel that doesn't require buying Google Cloud. Cerebras opened an order-of-magnitude inference gap on the workload pricing every contract right now. AMD moved from second source to second supply chain in a single week. Google chose mid-Missouri over its Northern Virginia footprint for the largest Midwest commitment of the year.
The financial models that underpin this entire industry are only as good as the assumptions they’re built on. Given how this week broke some of the big ones, we’re likely not far off seeing the consequences.
See you next week.