“Colossus was originally built to train Grok.” That sentence, from the most detailed reporting on SpaceX’s compute operation published this week, tells you almost everything you need to know about how the AI power landscape has shifted in the past 18 months. The facility that Elon Musk built to give xAI a competitive compute advantage has become the largest commercial AI compute landlord in the world. The competitors who were supposed to be disadvantaged by not having access to it are now paying him $80 billion to use it. This is not a normal market outcome. It is a structural fact about who controls the physical infrastructure of intelligence.”
Ground Truth · Episode 14 · June 25 2026SpaceX’s Colossus supercomputer facility in Memphis, Tennessee was announced in 2024 and built at extraordinary speed — 100,000 Nvidia H100 GPUs operational within approximately 120 days of breaking ground, in what Musk described as the fastest supercomputer buildout in history. The stated purpose was to give xAI the compute required to train Grok at frontier scale without depending on cloud providers owned by competitors. That strategic logic remains intact. What has changed is the scale and composition of who is using the facility. Colossus now has four major external compute tenants: Anthropic paying approximately $1.25 billion per month ($45 billion through mid-2029), Google paying approximately $920 million per month ($30 billion through 2029), Cursor (now SpaceX-owned following the $60 billion acquisition announced June 16th), and — announced this week — Reflection AI at $150 million per month for a $6.3 billion contract through the end of 2029. Combined committed external compute revenue: more than $80 billion. The facility built to train Grok is now the most commercially important AI infrastructure asset on earth, and the rent is being paid primarily by the companies it was designed to outcompete. Welcome to Episode 14 of Ground Truth.
$80 Billion in Committed Revenue. Four Tenants. One Landlord.
Anthropic: $45 billion, $1.25B/month, through mid-2029. The longest-established and largest Colossus contract. Anthropic uses Colossus 1 for Claude model inference and training. The contract was disclosed in Anthropic’s S-1 filing process and represents approximately 32% of Anthropic’s $47 billion revenue run rate flowing to a single vendor — one that also competes with Anthropic in the model market through Grok and that now owns Cursor, the coding interface that routes developer queries to Claude. The structural conflicts in this relationship are documented across Episodes 07, 10, and 11.
Google: $30 billion, $920M/month, through 2029. Colossus 2 — the second facility in the SpaceX Memphis campus — is Google’s compute location. Google is paying $920 million per month, making it the second-largest single tenant. Google also invests in Anthropic, runs its own frontier models competing with Grok and Claude, and now uses the same physical facility as its AI competitors. The Colossus campus in Memphis hosts the training and inference infrastructure for Anthropic, Google, and xAI simultaneously. These companies are competitors in the model market sharing compute landlord infrastructure.
Reflection AI: $6.3 billion, $150M/month, starting July 1 through 2029. The newest and most strategically interesting tenant. Reflection AI was founded in 2024 by Misha Laskin and Ioannis Antonoglou — both veterans of Google DeepMind, with Antonoglou having helped build AlphaGo. The company raised $2 billion in October 2025 at an $8 billion valuation and is now raising again at a $25 billion valuation. Its thesis: American governments, banks, defence contractors, and large enterprises want frontier-capable AI with open weights. They will not use closed US labs for sovereignty reasons — they need to be able to inspect and host the model internally. They will not use Chinese open-weight models for security reasons. Reflection is positioning as the third option: American, open-weight, and frontier-scale, with Pentagon and Department of Energy ties. No public frontier model has shipped yet. The compute contract with SpaceX gives them the infrastructure to build one. The compute contract at Colossus 2, where Google operates, means Google and Reflection AI train on the same physical infrastructure, with the same landlord, in the same Memphis facility.
Cursor (SpaceX-owned since June 16th). The fourth compute relationship is internal — Cursor’s model inference and training now runs on SpaceX infrastructure following the $60 billion acquisition. The consolidation is complete: SpaceX owns the model (Grok), the coding interface (Cursor), the compute (Colossus), and receives revenue from the three largest model providers in the world for using its facility to train their competing models.
Anthropic: $1.25B/month, Colossus 1, through mid-2029. Total committed: ~$45B. Google: $920M/month, Colossus 2, through 2029. Total committed: ~$30B. Reflection AI: $150M/month, Colossus 2, from July 1 2026, through 2029. Total committed: ~$6.3B. Cursor: internal (SpaceX-owned post-$60B acquisition). Total external committed revenue: $80B+. SpaceX build cost of Colossus: approximately $10 billion. ROI on external tenant revenue alone: 8x, before xAI’s own Grok training value. The facility built to give xAI a compute advantage has become a compute profit centre that funds SpaceX’s broader operations. The strategic and financial logic are now both fully aligned: Colossus is more valuable as a landlord than as an exclusive training asset.
The 2024 Nobel Laureate Who Built AlphaFold Just Left Google DeepMind for Anthropic.
On June 20th, John Jumper — co-lead of the AlphaFold 2 project at Google DeepMind and recipient of the 2024 Nobel Prize in Chemistry for that work — announced his departure from DeepMind after nearly nine years to join Anthropic. His role at Anthropic has not been officially announced, but the hire is widely expected to accelerate Anthropic’s AI for science programme, including its existing work on protein structure, drug discovery, and computational biology.
The AlphaFold context matters here. AlphaFold 2, released in 2020, solved the protein structure prediction problem that had been one of biology’s most difficult computational challenges for fifty years. It predicted the 3D structure of proteins from their amino acid sequences with accuracy that matched experimental methods — enabling drug discovery at a pace and scale that was previously impossible. Jumper and Hassabis shared the Nobel Prize for this. Jumper leaving DeepMind for Anthropic — which is simultaneously managing an IPO, recovering from the Fable 5 recall, opening its Seoul office, and signing $45 billion in compute contracts — is a statement about where the most consequential AI-for-science work is happening. It is also the second major DeepMind departure to Anthropic in a month, following Andrej Karpathy’s resignation from Anthropic (to be clear: Karpathy joined Anthropic earlier and resigned from it; he did not come from DeepMind). The talent signal from multiple simultaneous movements is notable.
Anthropic already operates Project Glasswing — the controlled-access programme for Claude Mythos Preview — which has used the model to find 23,019 vulnerabilities across 1,000-plus open-source projects. Jumper’s arrival positions Anthropic to extend that scientific application trajectory into biological and chemical domains where AlphaFold’s methods are directly relevant. A Nobel laureate in chemistry joining an AI company heading into an IPO is not a minor footnote. It is a statement about the frontier of where AI value creation is moving: from language and code toward physical science.
The man who won the Nobel Prize in Chemistry for using AI to solve one of biology’s hardest problems just joined the AI company whose flagship model is currently offline because the government pulled it thirteen days ago. Both of those things are true simultaneously. Anthropic is simultaneously at its most vulnerable and its most talent-attractive moment. That combination is either a sign of deep underlying strength or a very unusual inflection point. Possibly both.Neal Lloyd · Ground Truth, Episode 14
American, Open-Weight, Frontier-Scale. The Gap in the Market That $25 Billion Just Validated.
Reflection AI’s fundraising at a $25 billion valuation — without having shipped a public frontier model — tells you something important about the market gap it is addressing. The company’s thesis is simple and compelling: there is a large, well-funded segment of the AI customer market — governments, banks, defence contractors, critical infrastructure operators, large enterprises — that wants frontier-capable AI but cannot use the available options. Closed US labs like Anthropic and OpenAI raise sovereignty concerns because the models are black boxes that cannot be inspected or hosted internally. Chinese open-weight models like DeepSeek and Qwen raise security concerns regardless of their technical quality. The gap: American, open-weight, frontier-scale.
The Fable 5 recall — which demonstrated that the US government can pull access to a closed frontier model globally on hours’ notice — has made Reflection’s thesis considerably more compelling than it was on June 11th. An enterprise or government that experienced Fable 5 going dark on June 12th now has concrete evidence that depending on a closed model from a single US vendor is an availability risk with no backup. A model whose weights can be self-hosted, inspected, and operated independently of the originating lab’s decisions is a fundamentally different risk profile. Reflection AI is betting that a significant portion of the enterprise market will pay a premium for that risk profile. The $25 billion valuation implies the market agrees.
Reflection’s ties to the Department of Energy’s Genesis Mission and Pentagon AI programmes give it a government customer pipeline that neither Anthropic nor OpenAI has been able to fully develop — partly because both companies’ safety mandates create friction with military use cases that Reflection’s open-weight positioning avoids. The Fable 5 recall, in which the government demonstrated it does not fully trust Anthropic to manage its own model’s security, makes Reflection’s positioning more attractive to government buyers simultaneously. The company has not yet shipped a public model. When it does, the compute is already contracted at Colossus. The customer pipeline is already partially secured. The valuation at $25 billion before a public product implies that the market believes the product is a matter of when, not if.
The Highest Cybersecurity Benchmark Score Ever Recorded. And It Is Not Available to the Public.
On June 22nd, OpenAI launched the full version of GPT-5.5-Cyber as part of its expanded Daybreak cybersecurity initiative. The benchmark scores are striking: 85.6% on CyberGym (vs 81.8% for standard GPT-5.5), 39.5% on ExploitGym (vs 25.95%), and 69.8% on SEC-bench Pro (vs 63.1%). OpenAI calls the CyberGym score the highest ever recorded by any single model. The model is gated to vetted organisations through the Trusted Access for Cyber programme, which includes Akamai, Cisco, Cloudflare, CrowdStrike, Fortinet, Oracle, Palo Alto Networks, and Zscaler.
The context in which this launch lands makes it considerably more significant than a benchmark announcement. The Fable 5 recall was triggered, in part, by the government’s conclusion that Anthropic’s most powerful model posed cybersecurity risks it could not trust Anthropic to manage. OpenAI responded — within days — by launching a model specifically designed for cybersecurity applications, with higher capability benchmarks than anything previously available, gated through a government-connected vetted access programme that is the structural opposite of Anthropic’s open deployment approach. The competitive contrast is intentional and direct. OpenAI is positioning GPT-5.5-Cyber as the responsible alternative: powerful cybersecurity AI, available to trusted partners, through a controlled access programme that the government can monitor. It is the model Anthropic was accused of not managing securely enough, offered through a framework designed specifically to address those concerns.
SpaceX built a compute facility to train Grok. It now earns $80 billion from the companies that compete with Grok. A Nobel laureate joined the company whose flagship model the government recalled. A new AI lab raised $25 billion before shipping a product by addressing the market gap the recall created. And OpenAI launched the most powerful cybersecurity model ever built, gated to trusted partners, six days after Anthropic’s security practices were publicly questioned by the US government. The AI industry is not a market. It is a chess board, and every piece moved this week.Neal Lloyd · Ground Truth, Episode 14
Ground Truth, Episode 14 · June 25 2026
Neal Lloyd covers the real-world impact of AI — money, power, geopolitics, and the stories behind the headlines. Ground Truth is his daily AI news and analysis series on emdexter.blogspot.com.
- Ep 01The Gold Rush
- Ep 02ChatGPT Knows Everything
- Ep 03Siri Is Now Google
- Ep 04America’s AI Law Is a Mess
- Ep 05Is AI Taking Your Job?
- Ep 06Microsoft vs Everyone
- Ep 07SpaceX Is Trading
- Ep 08The Government Pulled Fable 5
- Ep 09Trump and Bernie Want to Own AI
- Ep 10SpaceX Buys Cursor for $60B
- Ep 11The Fable 5 Truth
- Ep 12The Week That Changed Everything
- Ep 13Bots Now Outnumber Humans
- Ep 14Colossus: $80B Compute LandlordYou are here
