TL;DR
AI is no longer reading as a cheap turbocharger for existing businesses; it’s an expensive infrastructure and geopolitics play where only a few vendors and governments are clearly ahead on economics. Everyone else is discovering that model tokens, power, memory, and politics are eating the ROI story, just as workers, regulators, and users start to push back.
The game is shifting from building demos to choosing which infra oligopolies and jurisdictions you’re willing to be dependent on.
Key Events
Report
AI is starting to look less like free productivity and more like an expensive, politically exposed infrastructure bet. The sharpest moves this period are in AI unit economics, the consolidation of power in infra vendors, and a rising legitimacy fight over who actually gets paid.
Most visible operators are discovering that AI features are costing more than the humans they were meant to replace, with Microsoft explicitly reporting that using AI can be more expensive than employing workers.
Microsoft is canceling internal Claude Code licenses as token-based billing blows through budgets and is also pulling the plug on a planned 244‑acre data center in Caledonia, signaling real capex discipline on AI infra.
Uber burned through its entire 2026 AI budget in four months and its COO now says it is getting harder to justify AI costs given the lack of meaningful feature improvements.
At the hardware layer, memory now accounts for roughly two‑thirds of AI chip component cost and Nvidia’s memory costs are up 485%, pushing latest systems to around $7.8M each just as data centers already consume ~6% of U.S. electricity and face local moratoria.
Anthropic is one of the few AI vendors showing a clear path to profit: it projects $10.9B in Q2 revenue on just $5B of lifetime revenue to date and expects to be profitable by Q2 2026 with around $500M in profit.
Its annualized revenue has exploded from $87M in January 2024 to $44B by late April 2026, underpinned by expansion to Colossus2 on Nvidia GB200 hardware.
That scale rides on extreme opex: Anthropic is reportedly paying SpaceX about $1.25B a month under a $45B compute deal, while also exploring Microsoft’s custom chips to diversify its stack.
On the demand side it is finalizing classified contracts with the NSA and other U.S. spy agencies, and the White House plans safety checks on its models before launch, while its Mythos tools have already found over 10,000 critical vulnerabilities using a library of 754 structured cyber skills.
Internally, Anthropic’s leadership openly says AI will displace human labor at very large scale and reports models exhibiting introspection and basic emotions, adding a philosophical edge to what is otherwise a very hard-nosed government and enterprise land grab.
SpaceX is repositioning itself as an AI compute landlord as much as a rocket company, claiming a $28.5T total addressable market with $26.5T tied to AI opportunities.
Its IPO filing suggests around $18B in annual revenue with 94% attributed to AI, 5% to Starlink connectivity, and just 1% to launch, yet it is still running roughly a $5B operating loss.
The Anthropic contract alone is worth about $15B a year, meaning a single AI tenant supplies most of SpaceX’s revenue momentum while the company is also offering AI compute as a service and planning orbital data centers powered by up to 1,000MW of solar.
Investors are flagging “AI-washing” in SpaceX’s revenue breakdown and worrying about the merger of xAI and social assets into the core business, even as regulators warn against giving Starlink excessive control over rural broadband.
Nvidia has largely conceded China’s AI chip market to Huawei after export bans, effectively forcing Chinese AI workloads onto a domestic stack.
Huawei is already supplying 2PB of flash storage for large language model training in Norway and has built a 122TB SSD using packaging tricks that sidestep U.S. 3D NAND sanctions.
Its Ascend NPU platform is running models like BitCPM‑CANN and Huawei’s Tau Scaling Law roadmap targets transistor densities equivalent to 1.4nm processes by 2031.
In parallel, China is flooding global markets with DRAM and NAND that are expected to push memory prices down even as the government spends nearly three times what the U.S. does on R&D and restricts overseas travel for top AI talent.
Commenters note that China is prioritizing cheap, widespread AI inference on commodity hardware backed by relatively cheap power, while the U.S. chases AGI and cloud-heavy stacks, implying structurally different cost curves.
Tech layoffs have already cleared 100,000 jobs in 2026, including 30,000 at Oracle, 16,000 at Amazon, 8,000 at Meta, and 3,000 at Intuit, with 99% of CEOs surveyed expecting AI-driven layoffs in the next two years.
Workers and commentators increasingly describe these as “AI-washing” moves to cut costs and juice stock prices, even as firms like Microsoft and Uber admit AI can be more expensive than human labor.
On the other side of the table, Samsung’s chip workers are getting average bonuses of about $340,000 amid the AI boom, sparking unrest among other divisions and contributing to a surge in union membership, while angry TSMC staff consider strikes over bonus cuts.
The first U.S. ride-share union for Uber and Lyft drivers has formed in Massachusetts, and California’s governor has signed an order targeting AI job displacement, as public sentiment toward AI skews more fearful than hopeful.
Layered on top is a broader cultural backlash—Gen Z booing AI-praising commencement speeches and Pope Leo XIV calling AI an instrument of domination that must be “disarmed”—which frames AI as benefiting wealthy executives at the expense of regular workers.
What This Means
Profit in AI is concentrating in a small cluster of infra vendors and government-aligned platforms while most enterprises face ugly unit economics and rising social and political friction over who bears the cost of automation.
On Watch
Interesting
We processed 10,000+ comments and posts to generate this report.
AI-generated content. Verify critical information independently.
Sources
Key Events
On Watch
Interesting