Morning briefing · 08:00
Power, chips and sandboxes: the foundations of the AI race
What AI companies compete over is expanding to the foundations beyond the model. That includes the power to run models, the chips to hold them, and the environments where agents execute code. Today's news included a report that a frontier model company bought power generation equipment. There was also news of a coding platform overhauling its execution isolation architecture, and a statement on the division of roles in the chip supply chain. All of these point in the same direction.
Model companies seeking direct control of power
The clearest signal came from power. The Financial Times reported that Anthropic secured gas power generation equipment for about $3.4 billion2. According to the report, the aim is to reduce uncertainty over data center power supply without relying on utilities. Until now, model companies have rented cloud capacity and grid power. This report shows that relationship is changing.
The change matters for another reason: power has become the variable that sets the ceiling on compute. With regulatory pressure on data centers growing, securing generation equipment in advance lets a company bring schedule risk in-house. However, the equipment's capacity, location and start-up date are not given in the source material, so it is too early to judge how much compute it will actually support. In Korea, XIIlab also decided to issue 10 billion won in convertible bonds for its GPU-based AI data center business11. The scale differs, but capital is moving toward infrastructure in both cases.
The same pressure on chips and execution environments
The same trend appears in the layers below power: chips and software execution environments. After reports that TSMC was considering 3D DRAM at NVIDIA's request, a TSMC spokesperson issued a statement6. It said the company has always focused on pure-play foundry and will remain a logic chip supplier. Analysis that the technology would take years to commercialize was also relayed. Taken together, the memory shortage is likely to remain a problem to be solved within the existing supply structure for the time being.
In automotive chips, an approach that consolidates functions on one chip emerged. AutoChips unveiled the AC8035, an SoC that handles the digital cockpit, L2/L2+ driver assistance and parking on a single chip and also runs on-device models (models that run directly on the device)9. Engineering samples are due in Q1 2027. It is an example of where and at what cost models run becoming part of product competitiveness.
At the software layer, Replit rebuilt its sandbox for agent code execution on microVMs7. Each sandbox now has its own kernel. eBPF (a technology for observing kernel activity) and Linux security modules monitor and block behavior. The company explained that even if the guest kernel is compromised, reaching the host requires crossing an additional virtualization boundary. In an era of agents running arbitrary code, the isolation architecture itself becomes the foundation that underpins trust in the product.
Variables that block efforts to secure foundations
Yet efforts to secure these foundations do not always go to plan. Chinese authorities barred SoftBank from acquiring the China operations of ABB's robotics division. According to Bloomberg, the two companies are discussing separating the China business into a standalone entity so the rest of the acquisition can proceed4. The total deal is worth $5.4 billion. The case shows that physical foundations can be split apart by regulation and regional borders.
Nor has competition at the product layer eased. Bloomberg's Mark Gurman reported that Apple CEO John Ternus is pushing to accelerate the timeline for turning Siri into a powerful AI agent3. OpenAI's Dots and Meta's Muse were cited as competitors. This means the competition users actually feel is still playing out in agent products. Securing foundations does not replace that competition. It is closer to a condition that supports it.
The takeaway from today's news is that model performance alone is no longer enough to judge the sustainability of an AI service. When choosing a vendor, practitioners should also consider how it sources power and chips and how it isolates agent execution. Going forward, watch whether the scale and location of Anthropic's power equipment are officially confirmed. Whether Chinese authorities accept SoftBank and ABB's plan to separate the China business is another point to check.