Susana SchwartzAnthropic's strategy: heterogeneity in compute and architecture
In yet another staggering infrastructure investment, Anthropic is hoping to power Claude’s exponential growth by investing $35 billion for long-term compute access through Nvidia-backed Lamda (with Nvidia itself holding the lease on the data center, as well as providing chips to Lamda). Hut 8 will develop the facility, which will deliver about 350 MWs of capacity, according to the Wall Street Journal.
Â
This is just the latests in a slew of deals Anthropic has forged with hyperscalers, neoclouds, AI accelerator manufacturers, and others. Anthropic is taking direct long-term leases on raw land and power assets, even repurposing Bitcoin mining companies in places like Kentucky and Texas.
Â
With a $45 billion deal struck not too long ago, Anthropic is poised to tap an Nscale microgrid campus in West Virginia – a 6-year reservation of Nvidia compute coming in at around 460 MWs.
Â
In Texas, Anthropic will be the anchor tenant in the Google-backed Nexus data center in Hubbard, Texas, which will bypass ERCOT by building a 1.6 GW natural-gas power plant for completely off-grid power. In addition, it also got a $19 billion compute lease at TeraWulf’s Justified Data Campus in Hawesville, Kentucky, where it will access a massive 400 MW cluster over a 20-year term.
Â
Like many of the leading AI companies, Anthropic is realizing that heterogeneity, as hard as it is to say, is the easiest way to procure enough compute. To give you an idea of the growing need for capacity, Anthropic in late 2025 operated on roughly 1.4 GW of capacity. By end of 2026, it will need roughly 5.0 GW.Â
Â
By designing the framework that compiles on Nvidia GPUs, Google TPUs, and Amazon Trainium chips, Anthropic marches toward the compute it needs, but avoids the risk of supply chain stalls it would possibly incur with a single manufacturer.
Â
And, by maintaining native hardware partnerships with the largest hyperscalers, Anthropic gains momentum with enterprises that want to avoid vendor lock-in, and potentially avoids the pitfalls of a more monolithic architecture.Â
Susana Schwartz
Technology Editor
RCRTech
AI Infrastructure Top Stories
Device-to-data center 6G connectivity: Qualcomm’s horizontal 6G strategy is steeped in diversification, with a pedigree of low-power, connected computing that is expanding from a wide range of devices into data centers, software and physical AI.Â
AT&T OTel telecom AI: For “right-sized hardware” in its data centers, AT&T has taken GSMA’s open-source AI model, OTel 2.0, and pushed it into full production deployment — with Dell and AMD supplying the hardware muscle.
AI Today: What You Need to Know
Verizon signs 2nd Google deal: Verizon’s dark-fiber agreement with Google is worth over $1 billion, with Verizon renting physical fiber to connect Google’s data centers – a new revenue stream that begins layering in during 2027.Â
London-Tokyo peering: LINX and JPIX announced new direct network routes between London and Tokyo, which will enable European networks to access one of Asia’s largest interconnection hubs with low latency and local peering opportunities.
‘The Barn’ logs 1 million craft hours: Leaders behind OpenAI’s data center project, as well as the Michigan Building and Construction Trades Council, celebrated the milestone, along with Oracle, Related Digital and Walbridge.
‘Data centers – a constant construction project’: Â Brittany Kaiser, CEO of Alpha Compute says data centers are driving electrical, concrete and plumbing jobs, with construction projected to reach $684 billion by 2031, says Merlo America.
Starcloud to test Fermi paradox: The founders of orbital data center start-up Starcloud will send a refrigerator-sized spacecraft toward Alpah Centauri – an 80,000-year journey that will cost about $15 million, to test the Fermi paradox.
RCR Events
Industry Resources
Report: NTN in motion — evolving standards, expanding services
