Newsletter · May 31, 2026
Weekly Digest 22
SoftBank commits up to €75B to a 5GW AI buildout in France, TSMC says energy efficiency is now the chip industry's binding constraint, and enterprises start trading tokenmaxxing for ROI discipline.
Topics we are tracking
SoftBank and France's AI infrastructure bid
Source: Financial Times — SoftBank plans up to 5GW data-center buildout in France
SoftBank announced one of the largest AI infrastructure commitments in Europe this week. The group plans to develop and operate up to 5GW of AI data-center capacity in France, representing up to €75B of investment. The first phase is a €45B buildout targeting 3.1GW of capacity in Hauts-de-France by 2031, with initial sites planned for Dunkirk, Bosquel, and Bouchain.
This is one of the first major announcements in Europe's AI infrastructure strategy. Europe is not going to outspend the US on frontier model companies or hyperscaler balance sheets. But France does have a real asset base: low-carbon baseload power, large industrial sites, a relatively centralised state, nuclear operating expertise, and a government willing to treat AI infrastructure as strategic industrial policy.
SoftBank is partnering with EDF on the Bouchain site, giving a former power-plant site a second life as AI infrastructure. Schneider Electric is also involved at the Port of Dunkirk, where it plans to build a large-scale industrial production cluster, including facilities for data-center enclosures and integrated power modules. That turns the project from pure compute capacity into a broader electrical-equipment and industrial-supply-chain story.
The read-through is clearest for Schneider Electric and the broader European electrical-equipment stack, where AI demand is pulling forward investment in power modules, cooling, switchgear, grid interfaces, and prefabricated data-center infrastructure. It is also positive for France's industrial-policy credibility. The country has managed to turn nuclear power, grid access, and industrial land into a plausible AI-infrastructure wedge.
The question is financing and execution. SoftBank has become one of the most aggressive balance sheets in AI, with exposure across OpenAI, Arm, Stargate-style infrastructure, robotics, and now European data centers. A 5GW buildout requires power contracts, grid upgrades, debt partners, equipment availability, permitting, customers, and years of construction discipline. The headline number is therefore less important than whether the first 3.1GW phase turns into real contracted capacity.
SoftBank's AI commitment spree
TSMC and the AI power wall
Source: Reuters — Energy use forcing a rethink of AI chip design, TSMC says
Source: SemiAnalysis — Inside the 800VDC revolution
TSMC said this week that energy efficiency has become the most important attribute its customers are asking for in next-generation chip technology. Kevin Zhang, TSMC's deputy co-COO and senior vice president, said the shift is visible across edge devices, mobile, IoT, high-performance computing, and data centers. Customers still want more compute, but they increasingly need it inside a tighter power envelope.
Large AI data centers need gigawatts of power. The chip industry therefore has to make each watt go further. TSMC is targeting roughly 30% efficiency improvement per generation and is preparing for a world where individual chips can exceed 1MW before the end of the decade. That pushes the roadmap toward better power delivery, more advanced packaging, integrated photonics, new thermal materials, and tighter memory integration.
The SemiAnalysis 800VDC piece shows the same pressure moving into the data center. As AI racks approach hundreds of kilowatts, the old power architecture starts to strain. Lower-voltage distribution means higher current, more copper, more conversion equipment, more heat, and less room for compute. SemiAnalysis estimates that moving toward 800VDC can cut facility-level power consumption by around 5%. At gigawatt scale, that becomes a meaningful amount of electricity and capex.
AI scaling is becoming a full-stack power problem. It starts at the transistor and package level, then moves through HBM, rack power shelves, busbars, cooling systems, substations, and grid connections. Nvidia, AMD, Google, Microsoft, and the other large AI silicon buyers need more peak performance, but the more relevant metric is increasingly useful compute per watt, per rack, per megawatt, and per dollar of data-center capex.
The foundry helps set the practical power envelope for future AI systems. It also expands the set of semiconductor bottlenecks investors need to track. Advanced packaging, HBM integration, optical interconnects, substrates, liquid cooling, power modules, and design-technology co-optimization are all becoming part of the AI scaling stack.
Tokenmaxxing meets ROI discipline
Source: Axios — AI spending and the enterprise ROI question
Source: The Information — Meta shutters internal AI token leaderboard
Source: arXiv — Token consumption in agentic coding tasks
Tokenmaxxing has been one of Silicon Valley's stranger AI adoption metrics over the last few weeks. The idea is simple: push employees to use more AI tokens, more often, across more workflows. Meta had an internal employee-built leaderboard tracking token usage in April. Amazon later shut down a similar Kiro leaderboard after employees started optimizing for usage rather than useful output.
The original logic was not completely irrational. Companies want employees to build fluency with AI tools, and usage is easier to measure than workflow redesign. Token leaderboards force experimentation. They make AI adoption visible. They also create social pressure inside engineering organizations that might otherwise treat AI tools as optional side projects.
The problem is that token usage is a weak proxy for productivity. Axios reported that corporate America is beginning to feel AI sticker shock. Microsoft has started scaling back Claude Code licenses in parts of the company. Uber's COO said the company has not yet seen a clear link between heavier AI usage and better product outcomes. The issue is not whether AI tools are useful. The issue is whether more inference spend reliably produces more output.
The strongest evidence comes from a recent paper on agentic coding tasks. The authors find that AI agents can consume roughly 1,000x more tokens than code chat or code reasoning tasks. Runs on the same task can vary by up to 30x in total token usage, and higher token consumption does not reliably translate into higher accuracy. Cost is therefore not just high; it is unpredictable.
That creates a different software cost structure from traditional SaaS. A normal SaaS seat has a relatively predictable marginal cost. An agent can call models repeatedly, inspect files, re-read context, run tools, recover from mistakes, and loop until it reaches a stopping condition. The bill depends on workflow design, context size, model choice, tool behavior, and how well the system knows when to stop.
This is why a new layer of AI cost tooling is starting to appear. Enterprises need token observability, workflow-level cost attribution, cheaper model routing, caching, context compaction, approval thresholds, and cost-per-outcome measurement.
OpenRouter raised a $113M Series B this week, with strategic investors including Alphabet's CapitalG, NVentures, ServiceNow, MongoDB, Snowflake, and Databricks. That is the right market signal. Enterprises will use more tokens, not fewer, but the tokens need to be allocated intelligently to the right model. A PhD-level model should not be used for a file-name change or a formatting task.
