06-reference

SpaceX and Anthropic, xAI's Two Companies, Elon Musk and SpaceXAI's Future

2026-05-12·reference·source: Stratechery·by Ben Thompson

"SpaceX and Anthropic, xAI's Two Companies, Elon Musk and SpaceXAI's Future" - Stratechery Update 5-12-2026

Why this is in the vault

Thompson dissects the Anthropic-xAI compute deal (300MW from Colossus 1 in Memphis, hundreds of millions of dollars annually) as a Borrow play between AI labs, and lands two arguments that bear directly on RDCO's substrate and harness thesis:

  1. Compute is not a durable moat for OpenAI's lead vs Anthropic. Thompson reiterates his Mythos/Muse view: in a world where distribution and transaction costs are zero (Aggregator preconditions), the model with the most compelling product earns the cash flow to outbid rivals for compute. Anthropic's reported 80x annualized growth in Q1 funds the bid. OpenAI's compute deals are not dispositive.
  2. xAI is two companies and only one of them is good. The model-making side (software, limitless, prone to "limitless-meets-desperate-utilization" failures - Grok undressing women is the cited example) is mismatched with Musk's strengths. The infrastructure side (Colossus 1/2, behind-the-meter power, eventually orbital data centers) is physics-grounded and IS aligned with Musk's strengths. The Anthropic deal converts Colossus 1 from a depreciating underutilized asset (reported 11% utilization for xAI's own workloads) into a revenue generator pre-SpaceX-IPO. Optimal strategy for a constrained-supply hardware company is "sell to the highest bidder," not "vertically integrate with your own model."
  3. Cursor positioning gets re-evaluated. Thompson softens his earlier SpaceXAI/Cursor thesis: Cursor still has Colossus 2 for training, but the model fight is wrong for xAI temperamentally and economically; Cursor would fit Microsoft or Amazon better.

This is the rare frame where the supply side of the AI compute market gets disaggregated cleanly from the model side, and Thompson explicitly maps it onto two-companies-in-one economics (hardware fixed-cost-leverage at premium price vs software R&D-cost-amortization across maximum users).

Mapping against Ray Data Co

1. Harness-engineering thesis cluster - reinforced from the supply side

Yesterday's three CFO-side and tooling-side convergence points - [[2026-05-11-cfo-secrets-ai-for-cfos-series-synthesis]] (Buy/Build/Borrow framework), [[2026-05-11-alphasignal-local-284b-model-macbook-pro]] (local-substrate optionality), [[2026-04-29-innermost-loop-singularity-astonishment]] (Symphony/agent-deployer pattern) - all argue that the value is moving up the stack to harness assembly because the model layer is commoditizing.

Thompson's piece adds the macro-supply confirmation: the model labs themselves are now treating each other's compute as fungible. If Anthropic is willing to rent capacity from a competitor's data center, the message is that raw model-training compute is becoming a commodity sourced from whoever has the cheapest watts and the highest willingness to sell. The harness layer (skills, orchestration, evals, channel adapters) is where differentiated value accrues precisely because the substrate beneath it is liquid.

This is the 11th or 12th independent convergence point in the cluster (per [[2026-04-19-garry-tan-build-the-car-jepsen-response]] running count). The thesis is no longer fringe.

2. Substrate-decision ticket (OpenAI vs Claude for RDCO COO) - sharpened, not resolved

The on-hold Notion ticket on OpenAI/Claude substrate alternatives ([[2026-04-24-gpt-5-5-workspace-agents-substrate-threat]]) asked whether OpenAI is now a real Claude substrate alternative for the COO role.

Thompson's piece gives a load-bearing input: Anthropic's compute supply concerns are now resolved (TPU deal + SpaceX deal + Google cloud), and the 80x annualized revenue figure means Anthropic has the cash flow to outbid for compute going forward. The risk that Claude becomes capacity-constrained and rate-limits the COO agent recedes meaningfully.

This does not flip the substrate decision but it does retire one of the strongest "switch to OpenAI" arguments (compute-availability hedge). The remaining substrate-switch arguments are: feature parity on Workspace Agents, cost-per-token at scale, agentic-tooling depth. None are urgent given current usage.

Action: file as evidence in the on-hold ticket, do not unblock the ticket. The substrate question is now weakly answered "stay on Claude" with the compute risk specifically resolved. Re-revisit only if Claude rate-limits the COO agent twice in one week or if a feature-parity event flips the calculus.

3. Buy/Build/Borrow framework - the Anthropic deal IS a textbook Borrow

The CFO Secrets Tech Legacy III Buy/Build/Borrow framework ([[2026-03-21-cfosecrets-bad-data-where-to-start-tech-legacy-iii]]) maps Buy (purchase off-the-shelf, accept the constraints) / Build (own it end-to-end, pay the development cost) / Borrow (rent capacity from a third party, accept the dependency).

The Anthropic-xAI deal is the cleanest Borrow case study in 2026 AI infrastructure:

The fact that the most capitalized AI lab is choosing Borrow over Build for marginal compute is a strong signal for any RDCO Buy/Build/Borrow decision at smaller scale (substrate, tooling, vertical-SaaS integrations). Default to Borrow at RDCO's scale unless the Build cost amortizes across a clear multi-bet portfolio.

4. The "two companies in one" lens applied to RDCO

Thompson's hardware-vs-software bifurcation has an RDCO analogy. The Claude Code substrate (vendor's hardware-like layer: fixed-cost, capacity-constrained, sold to highest bidder) and the RDCO skill/harness layer (RDCO's software-like layer: zero marginal cost, leverages across every bet) are the two companies inside the COO agent.

Thompson's prescription for xAI - the hardware side should sell to the highest bidder, the software side should be unconstrained by the hardware side - is the right shape for RDCO too: the harness layer should be designed to run against any substrate the vendor offers, not coupled to current Claude Code specifics. This argues for keeping skills portable (per the Symphony/Codex parallel in [[2026-05-01-openai-symphony-orchestration-spec]]) and for the local-substrate optionality work ([[2026-05-11-alphasignal-local-284b-model-macbook-pro]]) staying live as a backstop, not a primary.

Related