The AI infrastructure race is creating some unexpected alliances.
Anthropic just signed a major compute deal with SpaceX to lease the entire Colossus 1 supercluster in Memphis — over 300 MW of capacity with 220K+ Nvidia GPUs expected online within weeks.
A few interesting signals here:
• Claude usage limits are already increasing, including higher caps for Claude Code and fewer peak-hour restrictions. • Elon Musk is now effectively supplying compute infrastructure to one of OpenAI’s biggest competitors — despite publicly criticizing Anthropic only months ago. • Anthropic is also reportedly committing to a massive long-term cloud expansion with Google Cloud.
What stands out to me is how AI competition is shifting from just “models” to full-stack infrastructure strategy:
GPU supply Power availability Data center scale Cooling Energy partnerships Network capacity Capital access
We’re entering an era where compute itself becomes a strategic product.
This also reinforces a broader trend: companies that own infrastructure may end up with as much influence as companies building frontier models.
A new chapter in AI infrastructure may be unfolding far from land. Peter Thiel has led a $140M Series B investment in Panthalassa, an Oregon-based startup building autonomous, wave-powered floating compute platforms. The round reportedly values the company at close to $1B—signaling serious confidence in an unconventional idea: putting AI data centers in the ocean.
⚙️ How It Works
Panthalassa’s approach is equal parts engineering and environmental adaptation:
Each platform is an 85-meter steel node deployed in open ocean
Instead of traditional power sources, it converts wave motion into electricity
AI compute hardware onboard is naturally cooled by seawater, eliminating the need for energy-intensive cooling systems
The structures are self-steering, using hull design rather than engines to reposition in optimal waters
Connectivity is handled via SpaceX’s Starlink, transmitting AI outputs back to land-based systems
This is not just about floating infrastructure—it’s about decoupling compute from land constraints entirely.
🏗️ What Comes Next
The new funding will:
Complete a pilot manufacturing facility near Portland
Support deployment of the first wave-powered compute nodes in the Pacific
Target a commercial rollout by 2027
Thiel’s framing is bold—suggesting that compute infrastructure is entering a phase where “extraterrestrial solutions” are becoming viable. While space-based compute remains distant, the ocean offers a near-term, scalable frontier.
🌍 Why This Matters
AI infrastructure is hitting real-world limits:
Power consumption is skyrocketing
Cooling requirements are becoming unsustainable
Public resistance to large data centers is growing
Major players like Elon Musk and Google have explored futuristic alternatives—including space—but those remain long-term bets.
Panthalassa’s model sits in a practical middle ground:
Ocean = abundant energy + natural cooling
Offshore deployment = reduced regulatory friction
Mobility = dynamic optimization of compute locations
🧠 The Bigger Shift
This isn’t just a new type of data center—it’s a signal that AI infrastructure is becoming geographically fluid.
Instead of asking “Where can we build data centers?”, the question is shifting to:
“Where should compute live to maximize efficiency, cost, and sustainability?”
A new study out of Harvard University, published in Science, is raising serious questions about the future role of AI in clinical decision-making.
Researchers evaluated OpenAI o1-preview using 76 real emergency room (ER) cases—and the results weren’t subtle. The AI didn’t just perform well. It outperformed experienced physicians.
What the Study Tested
The study wasn’t theoretical or synthetic. It used:
Real ER patient cases
Raw electronic health record (EHR) text
Three stages of clinical decision-making
The AI had no special formatting, no structured prompts—just the same messy, real-world data clinicians deal with every day.
The Results: AI Took the Lead
At the initial ER triage stage, accuracy rates were:
67.1% — AI (o1-preview)
55.3% — Physician #1
50.0% — Physician #2
That’s not a marginal improvement—it’s a double-digit lead in diagnostic accuracy at the most critical early stage of care.
Even more interesting:
Independent physician reviewers could not distinguish between AI-generated and human diagnoses.
In other words, the AI didn’t just perform better—it blended in seamlessly with expert-level clinical reasoning.
A Real-World Moment That Stands Out
One case in particular highlights the potential impact:
The AI flagged a rare flesh-eating infection (necrotizing condition)
In a transplant patient
12–24 hours before the treating physician identified it
That kind of time advantage isn’t academic—it can be the difference between life and death.
What This Actually Means (And What It Doesn’t)
Let’s be clear: this does not mean AI is replacing doctors.
But it does signal something more practical—and arguably more powerful:
1. AI as a Second Set of Eyes
Doctors operate under pressure, fatigue, and time constraints. AI doesn’t. A system that consistently flags edge cases or rare conditions can act as a real-time diagnostic safety net.
2. Pattern Recognition at Scale
AI models trained across vast datasets can detect patterns that are:
Rare
Non-obvious
Easily missed in fast-paced environments like ERs
3. Decision Augmentation, Not Automation
The real value isn’t in replacing clinicians—it’s in augmenting their judgment, especially during:
Triage
Differential diagnosis
Risk identification
The Bigger Shift: AI Helping Doctors, Not Just Patients
Millions of people already use AI tools for personal health questions.
This study flips the narrative:
AI isn’t just for patients anymore—it’s becoming a tool for clinicians themselves.
And if a 2024-era model is already outperforming physicians in controlled settings, the trajectory is hard to ignore.
Where This Could Go Next
If integrated responsibly into clinical workflows, AI could:
Reduce diagnostic errors
Improve triage prioritization
Accelerate identification of rare conditions
Provide continuous clinical support in high-load environments
But this also raises real questions:
How do we validate and regulate these systems?
Who is accountable for AI-assisted decisions?
How do we integrate without over-reliance?
Final Thought
We’re not looking at a distant future scenario anymore.
We’re looking at a present-day signal:
AI is already capable of matching—and in some cases exceeding—human diagnostic performance in high-stakes environments.
The next phase isn’t about proving capability.
It’s about figuring out how to safely and effectively put that capability to work inside real healthcare systems.
A growing dispute between the White House and Anthropic is exposing a deeper issue in the AI race: who gets access to the most powerful models — and when.
At the center of the debate is Anthropic’s advanced AI system, Mythos, and a proposed expansion that would significantly increase private-sector access.
🔍 What’s Happening
Anthropic had plans to expand Mythos access from roughly 50 companies to nearly 120. On paper, it looks like a typical scale-up move. In practice, it triggered concern inside the U.S. government.
Officials pushed back, citing compute constraints — the fear that expanding access could strain infrastructure and limit availability for federal use, particularly in sensitive domains tied to defense and intelligence.
This friction comes as a new AI policy memo from the White House is being finalized — one that could reshape how agencies adopt and procure AI systems.
🧠 Policy Shift: Multi-Vendor AI Strategy
The upcoming memo is expected to encourage multi-vendor AI adoption across federal agencies, reducing reliance on any single provider.
This is a notable shift.
It also reportedly includes provisions that would allow agencies to bypass certain supply chain risk classifications, a move that could ease tensions with companies like Anthropic — even as legal and strategic disagreements continue.
In short: the government wants flexibility, redundancy, and leverage.
⚔️ Internal Friction in Washington
The situation isn’t just a government vs. company issue — there’s also disagreement within Washington.
Comments from figures like Pete Hegseth highlight a harder stance toward Anthropic, while others appear more focused on ensuring continued access to frontier AI capabilities.
This reflects a broader split:
One side prioritizes control, risk mitigation, and ideological scrutiny
The other prioritizes access, capability, and strategic advantage
🤖 The Bigger Picture: AI Parity Is Coming Fast
Adding urgency to the situation, models like GPT-5.5 are reportedly approaching similar cyber and reasoning capabilities as Mythos.
Former AI policy lead David Sacks suggested that most frontier models could reach comparable capability levels within six months.
If that timeline holds, exclusivity becomes temporary — and the battle shifts from who has access to how widely it’s deployed.
⚠️ Why It Matters
This isn’t just a policy disagreement — it’s a preview of how AI power will be managed:
Compute is now a strategic resource, not just a technical constraint
Access to frontier models is becoming a geopolitical lever
Government and private sector priorities are increasingly misaligned
The White House appears to be recalibrating — not necessarily backing away from Anthropic, but ensuring it doesn’t become dependent on any single player.
At the same time, internal divisions suggest that the U.S. is still figuring out how to balance innovation, control, and national security in the AI era.
If you zoom out, the signal is clear: AI isn’t just a technology race anymore — it’s an infrastructure, policy, and power struggle all at once.
