{"slug": "the-data-center-backlash-isnt-a-psyop", "title": "The data center backlash isn’t a psyop", "summary": "Chinese AI lab Z.ai reported its H1 2026 earnings, showing its model-as-a-service business grew 2,736% year-over-year, driving overall growth. The company's active user base reached 7.4 million, up from 5.8 million at the end of Q2 and a 144% increase since the start of the year, with paid daily active users growing 603%.", "body_md": "# The data center backlash isn’t a psyop\n\n## You can't simply blame China!\n\n**Tuesday. **[Oil prices are rising](https://www.wsj.com/economy/global/global-bond-yields-surge-as-oil-prices-fuel-inflation-worries-ef5c0192), stocks are down, and [the Fed may hike](https://www.cnbc.com/2026/09/01/fed-governor-barr-says-hell-support-rate-hike-if-inflation-doesnt-ease.html). I hope you have had your first coffee already. But don’t be blue; Amazon is having a worse week than your brokerage account. The FTC and 22 state AGs “[allege Amazon concealed unfair charges in its digital advertising auction pricing](https://www.ftc.gov/news-events/news/press-releases/2026/08/ftc-states-sue-amazon-over-secret-ad-surcharge-scheme).” When the inevitable legal fight wraps in 2049, it will be interesting to see who wins.\n\nToday we’re looking at Z.ai’s earnings, self-improving products, and why everything is China’s fault. So they tell me. **Your friend — Alex**\n\n*📈***Trending Up:**[Trump’s love of Putin](https://www.politico.com/news/2026/08/31/bessent-welcomes-russia-back-to-g20-as-europe-pushes-back-01058296)…[government borrowing costs](https://www.wsj.com/livecoverage/stock-market-today-dow-sp-500-nasdaq-09-01-2026?mod=hp_lead_pos1)…[Salesforce Ventures](https://www.bloomberg.com/news/articles/2026-09-01/salesforce-backs-hr-software-firm-hibob-at-3-2-billion-value)…[passing the buck](https://mediagazer.com/#a260831p18)…[self-driving taxis](https://www.cnbc.com/2026/09/01/waymo-and-zoox-expand-into-more-us-markets-as-robotaxi-race-heats-up.html)…[ACAB](https://techcrunch.com/2026/08/31/harvard-law-dropout-raises-6m-for-blue-voice-to-build-a-harvey-for-police-officers/)…[corruption](https://techcrunch.com/2026/08/31/polymarket-reportedly-raises-300-million-from-donald-trump-jr-s-investment-fund/)([remember this](https://www.cnbc.com/2026/08/27/trump-jr-told-gop-state-attorneys-general-dont-fight-prediction-markets-nyt-reports.html)?) …[a16z’s capital base](https://techcrunch.com/2026/08/31/a16z-brings-growth-fund-to-8-5b-days-after-launching-new-1-1b-fund/)…[what](https://www.reuters.com/world/middle-east/iran-urges-us-comply-with-interim-deal-after-trump-threatens-further-strikes-2026-09-01/)…[what](https://x.com/alex/status/2093861291823624347)…[📉](https://finance.yahoo.com/news/servicenow-pledges-1-5bn-investment-110000403.html)Trending Down:[Edtech in India](https://techcrunch.com/2026/09/01/indias-unacademy-sells-to-rival-upgrad-for-206m-about-94-less-than-its-peak-valuation/)…[Apple-OpenAI relations](https://techcrunch.com/2026/08/31/apple-shares-shocking-evidence-against-former-employee-accused-of-stealing-company-data-for-openai/)…[angering your compute partners](https://thenewstack.io/anthropic-spacex-cursor-compute-openai/)…[democracy](https://www.cnn.com/2026/09/01/politics/usps-whistleblower-mail-ballot-rules-trump)…[military preparedness](https://www.wsj.com/politics/national-security/army-secretary-resigns-after-months-of-friction-with-hegseth-9c124207)…[military preparedness](https://www.washingtonpost.com/world/2026/08/29/beijing-expands-purge-top-generals/)…\n\n**Z.ai earnings!**\n\nChinese AI lab Z.ai, of GLM model family fame, [reported its H1 2026 performance today](https://www1.hkexnews.hk/listedco/listconews/sehk/2026/0831/2026083101539.pdf). As with its [domestic rival MiniMax](https://www.cautiousoptimism.news/ais-30-trillion-sales-pitch/), Z.ai is growing quickly. Here’s its financial summary, converted into USD:\n\nDuring the first two quarters of 2026, Z.ai’s model-as-a-service (MaaS) business grew 2,736%, driving most of its aggregate growth compared to the same period in 2025. How did it post such exceptional expansion in what we’d call its core AI business? Massively expanding use. Per the company, its active user base soared to 7.4 million today, up from 5.8 million at the end of the second quarter and 144% since the start of the year. Since the start of the year, “paid daily active users [also grew] by 603%.”\n\nBut rising usage is only one vector that matters for Z.ai — it’s also raising prices, [as we’ve covered extensively](https://www.cautiousoptimism.news/american-ai-isnt-losing-the-price-war/). The company notes that it managed 40x token volume growth in its MaaS business since the start of the year, while its average “API selling price” rose by 101%. Or as the company trenchantly put it: “volumes surged while prices rose.”\n\nWhile Z.ai’s gross margin fell year over year, the difference came from its evolution from a majority on-prem company (selling models for local use) to offering them on a hosted basis. Thankfully, gross margins for its MaaS business rose from a small gross loss in H1 2025 (-0.4%) to a far healthier 24.6% in H1 2026.\n\nHow did Z.ai manage its greatly improved API economics? By building out a “cluster of over 100,000 domestic chips,” helping drive inference costs down “by 80% from the beginning of the year.” It gets more interesting: The AI lab claims that its “compute multiplier,” which measures “revenue generated from every RMB1 of computing power investment,” improved “by approximately 14 times year-on-year.”\n\nEfficiency gains at Z.ai are so rapid that the company now says that in “coding scenarios, the per-token inference cost on domestic chips has reached a level comparable in cost-performance to mainstream imported GPUs.” *Whoa.*\n\nFinally, GLM-5.3 Flash. Also known as Ox Alpha, 5.3-Flash is a model that Z.ai recently released (post-Q2) to great acclaim. Discussing the model’s early commercial performance, the company said that it drove 62 trillion tokens of inference in six days across its two partner surfaces (OpenCode, OpenRouter) while free, all while “deployed [on] a domestic chip cluster to serve ultra-large-scale real-world traffic.” The model remains [cheap and staggeringly popular](https://openrouter.ai/z-ai).\n\n**Takeaways:** Chinese AI labs are building competitive models. On domestic hardware. That serve a global audience. At increasingly efficient economics. Still, Z.ai remains deeply unprofitable, and small; $141 million in H1 2026 revenue is picayune compared to the figures generated in the American AI market.**Yes, but:** Z.ai has grown a*lot*since the end of the second quarter. The company said that its MaaS annual run rate reached $1.6 billion in August. Its Q3 results should dramatically outpace its second-quarter tallies.\n\n**The age of self-improving … product**\n\nForget self-improving AI, what if you built an agent that improved your product? Former unicorn, now public company Amplitude (product analytics root, various branches) pitched a new service called “Wave” during its [most recent earnings call](https://investors.amplitude.com/static-files/c501e2aa-38c5-4399-b395-0ef5f993325e). Calling it “the future of product development,” CEO Spenser Skates told investors that Wave builds “self-improving products that automatically recommend what to build next based on signals from your users.”\n\nAmplitude is [talking about Wave more now](https://amplitude.com/self-improving-products). I think the tool matters. Why? It shows what companies with lots of customer data can do with the resource. Wave builds on what its progenitor already collects, namely product usage data. With that information, it “decides what’s worth building,” constructs it, and offers up a pull request. You can either require Wave to wait for humans to review its work before something ships, allow for humans to step in when they want to, or simply let it rip.\n\n- Amplitude is\n[seeing gains](https://x.com/Jacob__Newman/status/2094473272796836097)from the product when deployed against its own offerings and digital footprint, including improved search, stronger onboarding completion, and increased engagement with blog posts. - We think a lot about self-improving AI models, but I think applied AI to create self-improving\n*other things*could matter quite a lot, too.\n\n**Did China do it?**\n\nThe data center debate has morphed from an intra-technology argument (are we building too much?) to a critical political issue (are you building too much near me?). The precise reasons for public discontent regarding data centers are manifold. I don’t want to [retread all our work on this issue](https://www.cautiousoptimism.news/all-that-data-center-agitation-is-working/), but secretive dealmaking, expensive subsidies, water and power usage, noise, and perceived bullying by companies looking to build new compute clusters *now* have each contributed to data centers’ [miserable polling](https://heatmap.news/daily/data-center-opposition-poll-collapse).\n\n**My priors:** I am pro data center, provided communities have a choice and aren’t forced to subsidize the installation. I don’t think local governance and an avoidance of corporate socialism are too controversial.\n\nIn recent months, a new meme has spread amongst technology types: China did it. The idea is that China is working to poison our national discourse regarding data center construction, allowing the rival economy to build out ample new compute while we remain mired in pre-AI FLOPs. There is some evidence for China working along these lines:\n\n### Subscribe to Cautious Optimism to unlock the rest.\n\nBecome a paying subscriber of Cautious Optimism to get access to this post and other subscriber-only content.\n\n[Subscribe](https://www.cautiousoptimism.news/subscribe/?utm_source=Cautious+Optimism&utm_medium=website&utm_campaign=article-paywall)", "url": "https://wpnews.pro/news/the-data-center-backlash-isnt-a-psyop", "canonical_source": "https://www.cautiousoptimism.news/the-data-center-backlash-isnt-a-psyop/", "published_at": "2026-09-01 14:47:15+00:00", "updated_at": "2026-09-01 14:55:40.537502+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-startups"], "entities": ["Z.ai", "MiniMax"], "alternates": {"html": "https://wpnews.pro/news/the-data-center-backlash-isnt-a-psyop", "markdown": "https://wpnews.pro/news/the-data-center-backlash-isnt-a-psyop.md", "text": "https://wpnews.pro/news/the-data-center-backlash-isnt-a-psyop.txt", "jsonld": "https://wpnews.pro/news/the-data-center-backlash-isnt-a-psyop.jsonld"}}