Anthropic buys 300 MW of compute from SpaceX — what the Neocloud movement means for Claude in the Mittelstand

What happened

On 20 May SpaceX published the financial details of the Anthropic compute contract announced on 6 May in its SEC IPO prospectus (Form S-1). Anthropic pays $1.25 billion per month through May 2029, with reduced rates in May and June during the ramp-up phase — around $40 billion in total, just under $15 billion per year. In return, Anthropic receives roughly 300 megawatts of GPU capacity on xAI's Colossus 1 in Memphis, Tennessee, a data center with more than 220,000 Nvidia GPUs (H100, H200, GB200). Anthropic co-founder Tom Brown simultaneously announced on X that the extension to Colossus 2 with GB200 capacity will ramp up from June. Either side can terminate the contract on 90 days' notice; in the S-1 SpaceX signals further similar contracts to come.

What it means

The more interesting statement here is structural, not financial. For the first time, a frontier lab is openly leasing compute from a direct competitor — and a compute owner with a growing capacity surplus (Grok app downloads fell to around 8.3 million in April 2026, down from more than 20 million in January) is now monetizing that surplus through its competitor. The industry already has a name for this pattern: „Neocloud“. AI labs become hybrid players who run their own inference and at the same time sell off excess capacity. With this deal, Anthropic operates a multi-vendor compute stack made of AWS, Google Cloud and Colossus — compute is decoupling structurally from the model provider.

What it means for the Mittelstand

For German Mittelstand companies, the deal shifts attention onto the compute path. Anyone running Claude in production through Amazon Bedrock, Google Vertex AI or directly against the Anthropic API now has a new actor in the processing chain: SpaceX as the contracting party, xAI Colossus as the physical GPU layer, Memphis, Tennessee.

Three questions become operational. First: which workloads run on Colossus, which on AWS or Google? Anthropic has not committed publicly; we read Tom Brown's reference to Claude Pro and Max subscribers as a signal that inference is in scope, not just training.

Second: the sub-processor list in your Anthropic DPA, or in the Bedrock/Vertex sub-processor appendices, has to reflect that SpaceX/xAI infrastructure can become part of the processing path. Check with your data-protection officer whether the record of processing and the DPIA need to be adjusted.

Third: third-country transfer to the US. The standard GDPR reflex is unchanged; the EU-US Data Privacy Framework was confirmed for now by the EU General Court's Latombe ruling of 3 September 2025, with the appeal before the CJEU still pending. In our assessment, the Musk corporate sphere as a sub-sub-processor adds a risk parameter to a DPIA — there is no regulatory codification of that assessment. For financial entities, DORA has been the primary framework for ICT third-party risk since 17 January 2025 (Art. 28 et seq.); BaFin withdrew the BAIT and VAIT circulars for DORA-subject institutions effective 16 January 2025, while MaRisk AT 9 and § 25b KWG remain in force as a complementary national layer. A change in the sub-sub-processor chain does not automatically become reportable — it becomes notifiable or reportable when the change affects a material outsourcing, or an ICT function supporting critical or important functions in the sense of DORA Art. 28.

What it means for the technical development

Architecturally, the deal confirms a trend that has been quietly visible for months: frontier labs are decoupling their compute supply from a single hyperscaler. Anthropic now runs a multi-vendor stack — AWS as the primary partner, Google Cloud with TPU, Colossus as the third pillar. For the platform layer, this means that the question „where does my inference run?“ is no longer answered implicitly by the model provider; the contract layer and the compute layer come apart.

Second: the Neocloud mode signals that GPU capacity is being handled as a tradable resource, not as a hyperscaler-exclusive asset. This plays into MCP and A2A architectures, because they abstract inference endpoints. If your agent architecture sits cleanly on MCP, you can swap the inference provider without rewriting the agent code — an inference-gateway layer becomes a production-relevant layer, not an exercise.

Concrete recommendation

In this order. First, take the sub-processor list from your Anthropic, Bedrock or Vertex contract and reconcile it against the 20 May disclosure; written confirmation of the processing path is a reasonable request to make of the provider. Second, work with your data-protection officer on whether your DPIA and your record of processing need to be extended to include the Colossus source. Third, financial entities check whether the change in ICT third-party risk under DORA Art. 28 et seq. crosses the materiality or critical-or-important threshold; MaRisk AT 9 and § 25b KWG add the national layer. Fourth, on the technical side: if your inference points directly at the Anthropic API today, this is the right moment for an MCP-capable inference-gateway layer that keeps inference providers swappable — Anthropic, Azure OpenAI, an open-weight model on-premises, depending on the sensitivity of the workload. The interesting question is not whether you keep using Claude. The question is whether you can say, for every production workload, which physical infrastructure it will run on tomorrow morning.

This post reflects our technical and strategic assessment. It is not legal advice and not a data-protection impact assessment.

Sources

• Anthropic — Higher limits for Claude and a compute deal with SpaceX (06/05/2026)
• TechCrunch — Anthropic will pay xAI $1.25B per month for compute (20/05/2026)
• Axios — Anthropic is paying SpaceX $15 billion per year (20/05/2026)
• BaFin — DORA arrives: changes to supervisory IT requirements (withdrawal of BAIT/VAIT/KAIT/ZAIT effective 16/01/2025) (09/01/2025)
• IAPP — European General Court dismisses Latombe challenge, upholds EU-US Data Privacy Framework (03/09/2025)