Here’s what’s on our plate today:

• 🧪 Who pays for the contest between Washington & Beijing?

• 📰 Microsoft turns on its partners, Zuckerberg's personal AI agents, and too much screen time hurts mood.

• 🛠️ Weekend To-Do: check your power exposure, read the AI Safety Index, try a free open model.

• 🗳️ Poll: who's actually paying the bill for the AI race?

Let’s dive in. No floaties needed…

The Laboratory

TL;DR

• Cold War arithmetic: superpower rivalry always generates a bill, and it lands on people who never agreed to race. Germany and Vietnam paid for the last one.

• Your power bill: U.S. electricity prices rose 6.9% in 2025, more than double inflation, as four tech giants spend $720B on data centers. Poorer households absorb the most.

• Safety, downgraded: Anthropic, OpenAI, Google DeepMind, and Meta have all weakened their pledges to pause development at dangerous capability levels. No lab scored above a C+.

• Nobody's the villain: slowing down means losing, so every rational actor cuts corners, and everyone ends up worse off.

• The stakes: when Washington and Beijing sit down in September, the room holds two delegations and nobody holding the invoice.

Who pays for the contest between Washington & Beijing?

During the Cold War, the now-dissolved Soviet Union and the United States were locked in a race to prove that their competing ideologies offered humanity's best path forward. Moscow believed communism was the future, while Washington saw capitalism as the engine of human progress. The two superpowers competed in nearly every sphere of human endeavor, from the space and nuclear arms races to technological supremacy. The rivalry accelerated innovation and produced remarkable scientific breakthroughs, but it also came at an enormous cost.

Much of that cost was paid by countries that belonged fully to neither camp. Germany remains one of the most striking examples, having been divided between the two superpowers for decades. History also reveals that the people living within the competing states paid a heavy price. In its effort to outpace the United States in the space race, the Soviet Union devoted enormous resources to technological competition, contributing to the economic strain that ultimately preceded its collapse. The United States, meanwhile, fought wars in countries such as Vietnam in its attempt to contain communism, leaving behind immense human suffering and destruction.

Today, Germany has long since been reunified, but the scars of that competition have not entirely faded. And now, another contest is unfolding between two powerful nations, one that could leave even deeper marks on humanity: the AI race between the United States and China.

The instinct, when this comparison comes up, is to reach for the most dramatic parallel: if the Cold War nearly ended in nuclear war, then the AI race must be headed for some equivalent catastrophe, a conflict over Taiwan or a machine no one can switch off. Those dangers deserve to be taken seriously, but waiting for the catastrophe means missing what the Cold War actually teaches, which is quieter and better documented. Great power competition always generates a bill; that bill arrives early, and it lands on people who were never asked whether they wanted to race. What was true in divided Berlin and in the villages of Vietnam is proving true again today, because the bill for the AI race is not a forecast waiting to be confirmed but an invoice already arriving, with line items that can be read.

The people missing from the conversation

The strangest episode of this summer is a good place to start reading them, because it shows who gets a voice in this race and who does not. On June 12, 2026, three days after the launch of Anthropic's newest AI models, the U.S. Commerce Department ordered the company to cut off access to non-U.S. citizens, citing national security concerns. Since an order like that cannot be enforced on a user-by-user basis, Anthropic switched the models off for the entire world, protesting that the government had never explained the threat, and for 18 days, one of the most capable AI systems ever built simply vanished from the market.

Then Washington reversed itself, and the models came back, after the company agreed to cooperate with officials on security standards. What forced the retreat was instructive: the loudest objection to the shutdown was that it amounted to a gift to China, handing its developers 18 free days to catch up, and when national security collided with the race, the race won. Weeks later, the two governments confirmed they would hold their first formal AI talks, in September. Every actor in that entire drama was a company or a state, and the people paying for the competition never appeared in it, because they appear somewhere else entirely: on utility bills.

The electricity bill comes due

The connection between an AI model and a household electricity bill runs through the data center, the warehouse of computers where these systems live. Four technology companies, Alphabet, Amazon, Meta, and Microsoft, are expected to spend roughly $720B this year, most of it on data centers, and those buildings draw power from the same grids that light ordinary homes, which means that when demand surges and supply cannot keep up, everyone on the grid pays the difference. U.S. electricity prices rose 6.9% in 2025, more than double overall inflation, and Goldman Sachs expects data centers to drive 40% of new demand for the rest of the decade. The bank's bluntest finding concerned who gets hurt most: poorer households, because electricity takes a larger share of a small budget than of a large one. By one estimate, $23B in higher power costs has already been shifted onto the public, and towns that once competed to attract data centers are now organizing to keep them out.

Racing away from safety

Money is only the most visible cost, because the race is also spending something harder to see: the safety margin the companies once promised to keep. A few years ago, the leading AI firms pledged to pause development if their own systems approached dangerous capability, and a safety review published this month found that Anthropic, OpenAI, Google DeepMind, and Meta have all weakened or abandoned those pledges, with some now saying a pause depends on what competitors do first. The reviewers called it moving the goalposts, and graded no company above a C+. The government followed the same logic when President Trump's June executive order cut the review period for new AI models from 90 days to 30, out of concern that a longer wait would put American firms at a disadvantage relative to China. None of this requires anyone to behave badly, and that is the most unsettling part. A study published this year laid out the trap in plain terms: when the prize for finishing first is large enough, cutting corners becomes the sensible choice for each racer, even though everyone ends up worse off when all of them do it. The same logic that once governed the missile silo now governs the release schedule.

The race gives as well as takes

An honest account has to admit that the race gives as well as takes. When Washington restricted chip sales to China, Chinese companies responded by giving their AI models away, free for anyone to download, and those free models have carried capable AI into countries that could never afford the American versions. Brookings finds that 24.7% of working people in wealthy countries now use AI tools against 14.1% in the developing world, a wide gap, but one that would be wider still without the giveaways the race produced. The cost side carries its own complications, since storm damage and the expense of replacing aging power lines are also pushing up bills, meaning data centers may be taking more blame than they have earned. Even the feared wave of job losses in the world's outsourcing hubs has not materialized, with unemployment in the Philippines falling to about 4% by March 2026, even as the country's call center industry quietly cut its long-term hiring forecasts.

The race enriches and extracts at the same time, and that is exactly what makes it so hard to stop, because every participant is behaving reasonably by its own lights. A company that slows down to test its models falls behind, a government that regulates falls behind, and a town that refuses a data center watches it get built one county over. Each decision is defensible on its own, and together they produce a bill that is automatically routed, without a signature, to whoever is not in the room. The difference from the last century lies in the address on the envelope. Where the Cold War pushed its costs outward, onto Korea, Vietnam, Afghanistan, and Angola, the AI race pushes them inward as much as outward: onto households in Virginia and Ohio, onto retirement savings tied to a construction boom that central bankers now list among the biggest risks to the financial system, onto workers whose displacement is arriving just slowly enough to be denied.

The world is beginning to notice the gap between who decides and who pays. A UN scientific panel warned at the first Global Dialogue on AI Governance in July that safety measures cannot keep pace with the technology, and a growing group of scholars argues that calling this an arms race at all makes everything worse, feeding nationalism and shrinking the room for cooperation. But a warning is not a mechanism, and neither the panel nor the scholars have any way to reach the people holding the invoice, and neither will the September talks. Whatever Washington and Beijing manage to agree on, the room will contain two delegations and no one else: no household whose power bill jumped, no engineer whose safety review got compressed, no call center worker watching the automation approach. The Cold War ended with a settlement between the two states that started it, and a much longer reckoning for everyone else. If this race ends the same way, the question it leaves behind is the one it began with: who agreed to pay for this, and when were they asked?

Friday Poll

Headlines You Actually Need

• Microsoft turns on its partners: Microsoft is now openly competing with OpenAI and Anthropic more than ever, signaling the era of cozy alliances is giving way to direct rivalry.

• Zuckerberg's personal AI agents: In Meta's Q2 2026 earnings, Mark Zuckerberg leaned hard into a vision of personal AI agents for everyone, framing them as the company's next platform.

• Too much screen time hurts mood: A new study links excessive time online to higher stress and worse mood, adding fresh evidence to the debate over technology's toll on wellbeing.

Weekend To-Do

• Check your own power exposure: Look up whether a data center is proposed near you via Data Center Watch, and see how it maps to local rate changes.

• Read the AI Safety Index: Skim the Future of Life Institute's AI Safety Index to see how the major labs graded on the pledges they made and walked back.

• Try a free open model: Download a capable open-weight model through Ollama to experience the "giveaway" side of the race the piece describes, capable AI at zero rental cost.

Meme Of The Day

The Toolkit

• Regie.ai: AI sales agent that researches prospects, writes personalized outbound at scale, and handles follow-ups so reps can focus on closing.

• Replit AI: A browser-based coding environment with an AI agent that builds, debugs, and deploys apps from prompts, with no local setup required.

• Sourcegraph: Code intelligence platform with Cody, an AI assistant that understands your entire codebase so it can answer questions and write code that actually fits.

Rate This Edition