China's Artificial Sun and America's New AI Job Title

Both countries are racing to secure what AI needs most: power and water. China is building new supply for both. The US is fighting over what it's already got.

China's "artificial sun," the EAST fusion reactor, hit a real milestone in June 2026: the largest fusion magnet ever built, a single coil weighing 582 tons, cleared testing. A new hydropower dam under construction in Tibet will use enough concrete to pave a two-lane highway around the Earth five times over, according to Bloomberg, and once finished it's designed to generate close to 300 terawatt-hours a year, roughly what the United Kingdom uses annually. On the water side, China activated the world's first commercial underwater data center off the coast of Shanghai in June 2025, a facility cooled entirely by seawater instead of freshwater, cutting the water draw that would normally come with running that much compute.

In the United States, both fights are happening over supply that's already stretched thin. US data centers now consume roughly 200 terawatt-hours of electricity a year, close to 6% of the country's total supply, according to a 2026 report from the International Data Center Authority. In Texas alone, existing data centers use an estimated 25 billion gallons of water a year, according to the Houston Advanced Research Center, which projects that figure could climb past 160 billion gallons by 2030. On July 18, 2026, opponents of data centers held 142 protests across 42 states in one coordinated day, according to Reuters, aimed at exactly this: the water and electricity data centers pull from grids and aquifers that were never sized for them. That pressure is already reshaping how US companies build, just not with concrete.

1. 1China: ~300 TWh/year once the dam is complete. Underwater data center cuts freshwater draw to near zero.
2. 2US: ~200 TWh/year already consumed, ~6% of the grid. ~25 billion gallons of water a year in Texas alone.

None of this is really about energy, not entirely anyway. It's a preview of where the real contest sits, one layer up from power grids and water pipelines. That layer doesn't play out the same way.

The frontier model race gets most of the attention: which lab ships the smartest system this quarter, whose benchmark chart climbs fastest. That race is real, but it doesn't decide whether AI does anything useful in someone's actual business or life. What decides that is everything that has to exist around a model before it does anything useful, and that part never shows up on a benchmark.

Where that "everything else" actually shows up depends entirely on which market you're standing in.

China Is Building the Physical Layer

In China, that layer shows up on the warehouse floor. SAP ran a pilot in November 2025 pairing its Extended Warehouse Management software with Unitree's G1 humanoid, the robot executing storage tasks guided in real time by SAP's system, with full SAP-level traceability on every move. RobotLAB sells access to robots like it the way you'd sell a SaaS subscription: Robot-as-a-Service, roughly $1,200 a month, no six-figure upfront cost, most deployments running within one or two business days of signing.

Figure AI's robots are still mostly a bet on what they'll eventually do, backed by a roughly $39 billion valuation as of September 2025. Unitree's are already on a warehouse floor with a monthly invoice attached.

It's backed by real capacity too: China added roughly 300 gigawatts of wind and solar capacity in 2025, a record, and solar overtook coal as the country's largest source of installed power in February 2026, according to the China Electricity Council, sooner than most forecasts expected.

The Software Layer Reverses the Picture

Claude Cowork, ChatGPT's Custom GPTs, and the wave of native connectors, skills, and plugins both companies shipped through 2026 exist because knowledge-worker productivity is where the commercial battle for US AI companies actually sits. Every one of those features lets someone extend the model into their own specific workflow.

Doubao, the app most Chinese consumers reach for first, isn't built that way. It's China's daily-active-user leader among AI apps, tied into Douyin (the Chinese version of TikTok), built around voice interaction and persona-driven conversation. It's excellent for consumers and, per one January 2026 review, lacks the deep reasoning capabilities required for complex enterprise tasks. The extensibility does exist in Chinese AI tooling: Kimi's developer CLI supports MCP and custom system prompts at the CLI level. It just doesn't show up in the mainstream consumer product, because that product was never built for the same job.

Same lesson, twice over: US companies responded to their market by going narrow, building deep into the one use case each could actually own.

US AI Companies Specialized Because the Market Forced Them To

The reshaping I mentioned earlier in this post shows up as specialization: ChatGPT owns the consumer layer. Anthropic built its commercial position around enterprise. Perplexity staked out research and search. None of that happened through coordination. There's no shared roadmap dividing up the market. Each company operates market-first and survival-first, in a landscape with no coordinated path forward, and specialization is what a company does when it needs a defensible piece of ground and nobody's going to hand it to them.

The forward deployed engineer role makes the same instinct visible at the level of individual hiring: OpenAI, Anthropic, and Google Cloud are all now hiring for it, embedding engineers directly inside client organizations to build the specific thing that client needs. Postings for the role grew more than 800% year over year by early 2026, according to SkillScouter.

That's not just something happening to companies from the outside. It shows up from the inside too, in the middle of an ordinary workday.

A friend of mine who works at a European marketing agency's China office has been living this directly. Her agency's French headquarters signed a partnership with Anthropic and handed her office a token budget to work with, generous on paper, tight in practice. She burns through it fast enough that she's now weighing a personal account just to cover the gap.

The contract solved the procurement problem. It didn't solve the deployment problem. Those are two different layers, and when a real workload outgrows a token budget, the question of whether a different model could do the same job for less money stops being theoretical.

Kimi K3 Closed That Question

Moonshot AI released Kimi K3 in July 2026, 2.8 trillion parameters, an absurd number even by 2026 standards, and it landed fourth on the Artificial Analysis Intelligence Index, the closest an open-weight model has ever gotten to the closed frontier. It leads the Frontend Code Arena leaderboard outright, ahead of every closed model tested, including Anthropic's own.

Whatever distance still separated open-weight models from the closed frontier narrowed enough in one release that it stopped being the safe assumption to lean on. Which is where my own work actually sits.

TikSense runs models from both US and Chinese labs inside the same production system. Running them side by side comes with real technical friction these days that I won't get into, but which lab's model runs a given track was never the hard decision. It came down to practical fit at the time, not a capability contest between two ecosystems, and that's exactly the point.

Model selection for each track took an afternoon. Getting the outputs to actually cohere into one trustworthy verdict took months. The model is infrastructure. The system around it is the product, the same argument I made building The Seven Dimensions of Enterprise AI, and the reason model selection was never where the real work happened here either.

China is building its layer in warehouses, with signed pilots and dedicated power to keep the compute running. The US is building its layer in software, connector by connector and engineer by engineer, embedded inside the exact companies that need it. Different layers, same instinct: figure out what has to exist around the model to make it actually work, and build that instead of waiting for the next benchmark to settle the argument for you.

This lives wherever you already are, Substack, X, or LinkedIn, drawn from building TikSense.