A few months ago, the Trump administration was publicly calling Anthropic a radical left menace to national security. Now its CEO is reportedly walking into the White House for meetings. That is a remarkable turnaround, and it seems to have a lot to do with one very well-timed product launch.
The backstory: Anthropic's relationship with the Pentagon collapsed fast in late February after the company refused to cross two specific lines — enabling domestic mass surveillance and building fully autonomous lethal weapons with no human oversight. Before that falling out, Anthropic had been deeply embedded in the US defense apparatus. Its models were the first commercially developed AI to be cleared for use on classified military networks. Losing that relationship was a big deal for both sides.
What followed was ugly. The administration publicly branded Anthropic a national security risk. Anthropic filed a lawsuit contesting that designation and won a temporary injunction. For weeks, the two sides seemed locked in a standoff with no clear off-ramp.
Then came Claude Mythos Preview. Anthropic's new cybersecurity-focused model is being positioned as a tool for finding dangerous vulnerabilities in the software infrastructure that underpins the modern internet — think major web browsers, operating systems, and financial platforms. Apple, Nvidia, and JPMorgan Chase have already signed on for access. The pitch is straightforward: find the holes before the bad actors do. Reports suggest the model's release prompted emergency conversations between major US bank executives and Federal Reserve Chairman Jerome Powell, which tells you something about how seriously people are taking its capabilities.
Anthropologic moved quickly to make sure Washington noticed. The company briefed senior US government officials on Mythos and its offensive and defensive cyber potential before the public announcement. Dianne Penn, a product management lead at Anthropic, confirmed those briefings happened, though the company declined to name who specifically was in the room.
The company also recently retained Ballard Partners, a lobbying firm with close ties to the Trump orbit, which signals that Anthropic is no longer content to win arguments in court and lose them in the hallways of power.
None of this means the underlying tension is gone. Anthropic has not walked back its red lines on autonomous weapons or surveillance. Those positions remain firm, and they remain the original source of the conflict. What Mythos Preview has done is give both sides a reason to talk again — a shared interest in cybersecurity that sidesteps the most combustible disagreements.
The fact that Dario Amodei was reportedly at the White House on Friday suggests that reason may be enough, at least for now. Anthropic did not confirm or comment on that meeting, which is itself a kind of answer.
The broader lesson here is one the rest of the AI industry is probably watching closely. When your government relationships sour, the fastest path back may not be through the courts or through public statements — it might be through launching a product that makes you too useful to stay angry at.
AI
40 Percent of US Data Centers Face Major Construction Delays in 2026
Here is a number worth sitting with: nearly four out of every ten US data center projects scheduled for completion this year are likely to miss their deadlines by more than three months. Not because of funding problems or strategic pivots, but because the physical world is not keeping up with the ambitions of Silicon Valley's spreadsheets.
The finding comes from an analysis by the Financial Times, which used satellite imagery from geospatial analytics firm SynMax to track how much actual ground-clearing and foundation work had been completed at project sites. That data was cross-referenced against permit filings and public statements compiled by industry research group IIR Energy. The result is one of the more concrete pictures we have had of where the AI infrastructure buildout actually stands versus where companies said it would be.
The culprits are familiar but worth spelling out. There are not enough skilled tradespeople — electricians, pipe fitters, the people who actually wire and plumb a massive facility — to staff multiple large-scale construction projects simultaneously. Executives working on OpenAI projects specifically flagged this shortage in interviews with the Financial Times. When you are trying to build data centers in multiple states at once, you are competing with yourself for the same limited pool of workers.
Power is the other wall everyone keeps running into. These facilities consume electricity at a scale that strains regional grids, and utility companies are not expanding infrastructure fast enough to meet demand. Tariffs on Chinese-made equipment, including transformers, have added cost and delay to a supply chain that was already stretched thin.
The improvisation required to keep projects moving is striking. Some developers are hauling in mobile gas generators on semi-trucks. Others are using turbine engines originally built for aircraft and warships. The fact that companies are retrofitting military-grade hardware to power consumer AI products tells you something about how strained conventional energy infrastructure has become.
Then there is the community resistance angle, which tends to get less attention but is becoming harder to ignore. Virginia, which hosts more data center capacity than anywhere else in the country, has seen public opinion shift noticeably against further development. A recent poll found most Virginians worried about land use, environmental impact, and the effect on household electricity bills. That last concern is well-founded — data center demand has a documented history of pushing utility companies to raise rates for everyone in a region, not just the tech tenants.
The Trump administration announced a Ratepayer Protection Pledge in March, with several major tech companies signed on. But the agreement has no legal teeth and no clear implementation mechanism, which means it is more of a PR gesture than a policy solution.
All of this matters because the AI industry's near-term ambitions are directly tied to how fast this infrastructure gets built. Delays do not just push back ribbon-cutting ceremonies — they push back the compute capacity that AI companies are counting on to train and run the next generation of models.
Source: Ars Technica
The finding comes from an analysis by the Financial Times, which used satellite imagery from geospatial analytics firm SynMax to track how much actual ground-clearing and foundation work had been completed at project sites. That data was cross-referenced against permit filings and public statements compiled by industry research group IIR Energy. The result is one of the more concrete pictures we have had of where the AI infrastructure buildout actually stands versus where companies said it would be.
The culprits are familiar but worth spelling out. There are not enough skilled tradespeople — electricians, pipe fitters, the people who actually wire and plumb a massive facility — to staff multiple large-scale construction projects simultaneously. Executives working on OpenAI projects specifically flagged this shortage in interviews with the Financial Times. When you are trying to build data centers in multiple states at once, you are competing with yourself for the same limited pool of workers.
Power is the other wall everyone keeps running into. These facilities consume electricity at a scale that strains regional grids, and utility companies are not expanding infrastructure fast enough to meet demand. Tariffs on Chinese-made equipment, including transformers, have added cost and delay to a supply chain that was already stretched thin.
The improvisation required to keep projects moving is striking. Some developers are hauling in mobile gas generators on semi-trucks. Others are using turbine engines originally built for aircraft and warships. The fact that companies are retrofitting military-grade hardware to power consumer AI products tells you something about how strained conventional energy infrastructure has become.
Then there is the community resistance angle, which tends to get less attention but is becoming harder to ignore. Virginia, which hosts more data center capacity than anywhere else in the country, has seen public opinion shift noticeably against further development. A recent poll found most Virginians worried about land use, environmental impact, and the effect on household electricity bills. That last concern is well-founded — data center demand has a documented history of pushing utility companies to raise rates for everyone in a region, not just the tech tenants.
The Trump administration announced a Ratepayer Protection Pledge in March, with several major tech companies signed on. But the agreement has no legal teeth and no clear implementation mechanism, which means it is more of a PR gesture than a policy solution.
All of this matters because the AI industry's near-term ambitions are directly tied to how fast this infrastructure gets built. Delays do not just push back ribbon-cutting ceremonies — they push back the compute capacity that AI companies are counting on to train and run the next generation of models.
Enjoyed this?
Get stories like this delivered every Tuesday — free.