“Frontier velocity can slow while deployment velocity accelerates.”
The most consequential number of the morning is not the Federal Reserve's expected 25-basis-point increase. It is 2.16 gigawatts.
Anthropic has signed its first Australian data-center lease, covering capacity at a planned 2.16-GW campus about 250 kilometers from Brisbane, according to Reuters. The project is expected to begin coming online in 2027. Crucially, the facility is intended for inference rather than training.
That announcement arrives days after Anthropic CEO Dario Amodei called for a slower pace of frontier-model capability development—an idea supported by OpenAI's Sam Altman and xAI's Elon Musk. This morning European Commission President Ursula von der Leyen backed efforts to "pace the frontier" and said she would convene leading AI laboratories. Meta CEO Mark Zuckerberg, however, explicitly rejected the need for coordinated slowing, arguing that competition and liability already give companies sufficient incentives to build safely.
At exactly the same time, the physical AI economy continues to expand. The deeper signal is emerging clearly: frontier velocity can slow while deployment velocity accelerates. That may define the next phase of the AI Supercycle.
The Fundamental Question
Building smarter vs. industrializing intelligence
Five Signals Shaping the Next Economy
Where the Supercycle is rotating
-
Anthropic's Australian agreement covers capacity at a planned 2.16-GW campus near Brisbane, expected to start operating in 2027 using renewable power and closed-loop air cooling. Reuters reports the capacity would be used for inference, not training; the project still requires approval from Australia's Foreign Investment Review Board. Economists cited by Reuters estimate Australian data-center investment could reach A$150 billion by 2030.
The causality chain may become: frontier models → enterprise adoption → agentic AI → billions of inferences → accelerated computing → data centers → electricity. Training is concentrated among a handful of laboratories; inference propagates through the entire economy. The largest opportunity may lie not in producing the next model but in producing the lowest-cost reliable intelligence per transaction. signal: Cost per token at industrial scale becomes the metric that matters.
-
SK Hynix is discussing a potential agreement with Intel that could involve manufacturing memory chips in the United States for the first time—one option leasing part of Intel's long-delayed Ohio facility, another a venture linking SK Hynix, Intel and major cloud companies. Talks are exploratory. SK Hynix, the leading supplier of high-bandwidth memory, closed 4.1% higher in Seoul; Intel rose 5.2% pre-market.
The bottleneck chain has become GPU → packaging → HBM → fab location → national security. Once HBM becomes scarce enough, the location of memory production becomes strategic. The question is no longer who designs the best accelerator, but who controls the full physical stack required to feed it. signal: Accelerated-computing sovereignty is spreading upstream through the semiconductor chain.
-
The U.S. 10-year yield is around 5%, after reaching 5.041%, its highest since 2007. Markets price roughly 92.5% odds of a quarter-point hike tonight, taking the target range to 3.75%–4.00%. Chair Kevin Warsh speaks after the decision, with investors focused on whether this is a one-off response to inflation or the beginning of renewed tightening.
Today's decision sets the shadow price of AI capital: AI expected return − risk-free rate − energy cost − security cost − execution risk = velocity spread. At 5%, an infrastructure project cannot justify itself by pointing to explosive demand alone—it must demonstrate utilization, contracted customers, power access, capital efficiency and free cash flow. signal: The split deepens between AI capacity and AI capacity that earns its cost of capital.
-
Europe plans seven large AI computing hubs, with 18 EU governments bidding to host them and collectively pledging around €3 billion in future compute purchases. By comparison, six largely U.S.-based hyperscalers are expected to spend more than $1.3 trillion by 2027. Large EU industrial consumers paid roughly $107/MWh in 2025—more than double the U.S. level. About 40% of European distribution grids are over 40 years old, €584 billion of grid investment is needed by 2030, and network connection can take as long as seven years versus roughly two to build a data center.
Of 68 major EU data-center projects since 2024, 43 are concentrated in France, Spain, Finland and Sweden. A 100-MW facility incurs roughly €254 million in annual electricity costs in Germany, €153 million in Spain and only €91 million in Finland—explaining Google's €13 billion Finland investment. signal: Europe does not merely have an AI funding problem; it has an electron-allocation problem.
-
Brent has eased about 0.6% today toward $108 per barrel after Saudi Arabia offered additional crude via Oman and U.S. data showed a 7.1-million-barrel inventory build. But refined-product markets remain exceptionally tight: Asian 10-ppm diesel refining margins have exceeded $87 per barrel, versus roughly $22 before the war, with near-term diesel spreads above $11.
The physical-economy shock travels through crude → refining → diesel → freight → logistics → food → industrial costs → inflation. A temporary easing in crude does not eliminate inflation risk if refined-product scarcity remains extreme—and the AI Supercycle now competes for physical capital in an economy where construction, grid equipment and backup generation are becoming more expensive. signal: Intelligence may be deflationary; the build-out required to industrialize it is not yet deflationary.
The Hidden Patterns Beyond the Headlines
Five structural shifts
The Model Race and the Deployment Race Are Separating
Anthropic can simultaneously argue for slower recursive model improvement and reserve capacity inside a 2.16-GW inference campus. Frontier training determines maximum capability; inference determines how widely it propagates. The next phase can look like slower capability velocity, faster propagation velocity. The market spent three years valuing the creation of intelligence; the Next Economy may value its propagation.
Inference Could Be More Physically Distributed Than Training
Training economics favor concentration—gigantic clusters, maximum accelerator density, massive power blocks. Inference economics differ: latency, sovereignty, data localization, reliability and enterprise proximity all matter. The architecture could migrate from a few superclusters toward a global network of industrial AI factories, greatly increasing the number of countries able to participate economically.
AI Safety Is Becoming Another Great Split
Europe has moved toward Amodei's position, with von der Leyen planning to pace the frontier. Meta has moved the opposite way, with Zuckerberg arguing competition and liability suffice and President Trump warning excessive caution could benefit China. The split runs through the industry itself: coordinated pacing versus competitive self-governance. Governance has become an economic variable.
Memory Is Moving From Component to Geopolitical Infrastructure
The possible SK Hynix–Intel partnership exposes a broader pattern: when a component becomes indispensable enough, governments stop treating it as a normal traded good. Oil demonstrated this; semiconductors demonstrated it; now HBM follows. The chain runs AI demand → HBM scarcity → strategic dependence → localization pressure → national capital → new industrial geography.
Europe's Real AI Gigafactory May Have to Be the Grid
Europe debates how many gigafactories to build; the deeper issue is whether its electricity system can support them. A €584 billion grid-modernization requirement is not adjacent to AI policy—it is AI policy. A continental network moving abundant nuclear, hydro, wind and solar toward industrial demand may generate more European AI capacity than another isolated subsidy.
What to Watch This Afternoon
Two catalysts and four gauges
The rotation down the stack
- Anthropic's contradiction is only apparent: slow the frontier, build 2.16 gigawatts—two actions that reveal a Supercycle moving from creating intelligence to deploying it.
- The Great Split moves again: not AI vs. non-AI, nor acceleration vs. slowdown, but intelligence creation vs. intelligence propagation.
- Models can slow while inference accelerates; regulation can tighten while deployment expands; capital can grow more expensive while high-return AI infrastructure still attracts investment.
- Europe can regulate frontier intelligence successfully yet still lose the economic race if it cannot produce the electrons required to deploy it.
- Today's equation: accelerated computing × inference × energy × capital × governance = industrial intelligence.