Anthropic is said to be in talks for a roughly $6 billion acquisition of Decart to extract more performance from its existing compute, while CoreWeave warns that diversifying beyond Nvidia carries cost and delay, and Cerebras shares slid on soft growth.
Strip away the model-launch noise and the day's real story is financial: the frontier is now gated by the price and useful life of silicon, not by algorithms. When a lab would rather spend billions on an efficiency layer than on more chips, it is telling the market that raw capacity is no longer the constraint, utilization is. Nvidia's parallel move to court financiers and shield the residual value of aging GPUs confirms the anxiety underneath the boom, that a two-to-three-year depreciation clock on hardware bought at cycle-peak prices could turn today's capex heroes into tomorrow's write-downs. CoreWeave's candor about the switching cost of leaving Nvidia is the tell: lock-in is now a balance-sheet feature, and the entire buildout is being underwritten on the assumption that inference demand compounds faster than the assets decay.
For investors, the divergence between Databricks-style software multiples and Cerebras-style hardware skepticism marks the line where capital is starting to discriminate between businesses that own depreciating iron and those that monetize its output. Efficiency, not scale, is becoming the moat.
For Japan, this reframes the AI conversation away from sovereign-model ambition toward compute discipline. Japanese enterprises and SIers such as the NTT Data, Fujitsu and NEC tier are largely renting frontier capability through hyperscalers, which means their AI cost base inherits every swing in GPU pricing, memory scarcity and yen-denominated import inflation. The strategic edge here is not building models but architecting for token efficiency, routing cheap fast tiers for routine work and reserving frontier calls for high-value tasks. SIers that can package this cost governance into managed services will out-earn those simply reselling API access.
The RPA and local dev implication is sharper still. As inference gets faster and cheaper, brittle rule-based automation loses ground to agentic workflows, and vendors like the UiPath-and-WinActor installed base in Japanese back offices face a migration window. The winners will be teams that treat compute as a metered utility to be optimized, not an unlimited resource to be consumed.
