Economics of the AI Build-Out
Research Rundown #150, plus new memos on Island, Astronomer, and more.
Research Rundown
In July 2025, The Wall Street Journal reported that AI capital expenditures, like data center hardware and infrastructure, have contributed more to 2025 US GDP growth than all consumer spending combined. But can this infrastructure-heavy model ever support the economics of AI?
As investors become increasingly concerned about the possibility of a bubble in both private and public market AI bets, critics of the economics of AI have been quick to remind everyone that no frontier AI company relying on third-party data centers is yet profitable. Defenders of AI's potential, on the other hand, point to the plummeting costs of inference. But model training sets have grown 350,000x in the last decade compared to only a 100x decrease in compute costs.
Meanwhile, from Jensen Huang to Sam Altman, AI leaders seem to be in competition about who can say the biggest number. $1 trillion? $4 trillion? But the fundamental question remains: will that level of investment ever result in profitability for AI companies?
For our exploration of that question, you can check out our full deep dive here.
Island is quietly transforming the most basic tool in tech—the browser—into an enterprise security powerhouse. Read our memo here for everything you need to know.
How Astronomer is turning the chaos of modern data pipelines into a competitive advantage. We break it down in our memo here.
Could Antora Energy’s thermal batteries be the missing link that makes clean energy truly limitless? Check out the full memo here.
What happens when AI stops just spotting anomalies and starts understanding the patterns behind them? That’s Goodfire’s bet. Unpack the full story in our memo here.
Rad AI is tackling one of healthcare’s biggest choke points: the flood of medical imaging. Get the details in our memo here.
JPMorgan analysts say they see no signs of excess in GenAI spending, with top cloud providers projected to fund at least 15% capex growth in 2026 from strong free cash flow. Private AI labs, sovereign funds, and new entrants like CoreWeave are also adding undeployed capital. Reasoning model adoption is just inflecting, agentic AI remains early, and China’s cloud spend is only beginning, suggesting sustained demand growth into 2027–28.
Anthropic has settled a lawsuit from authors who alleged the company illegally copied their books to train AI models. The deal, reached through mediation, avoids a trial that could’ve resulted in major damages of up to $150K per work. A June ruling had found Anthropic on solid legal ground for training under fair use but left liability over downloading seven million books unresolved. Settlement details remain undisclosed, though lawyers called it “historic” for class members.
China is outpacing the US in open-source AI, with its models now leading global benchmarks and adoption; around 80% of US AI startups use Chinese models, per a16z’s Martin Casado. The Economist notes top open models are all Chinese, narrowing the gap with proprietary systems and challenging America’s closed-source approach.
Crusoe, developer of OpenAI’s Stargate data center, is raising at least $1 billion at a $10 billion valuation; over 3× last year’s. Revenue is projected to climb from $250 million in 2024 to $2.2 billion in 2026, with cloud rentals reaching $18 billion by 2030. Expanding sites in Texas and Wyoming, it aims to undercut AWS, Oracle, and CoreWeave through vertical integration and power deals, and recently bought Tel Aviv–based Atero to improve GPU efficiency.
Some argue that MIT’s claim that 95% of enterprise AI projects fail reflects a long IT pattern: only 9% of large projects succeeded in 1994, rising to just 10% by 2020. Failures stem from turning business needs into specs, rollouts, and adoption, now worsened by overhyped models and thinly resourced mandates. The study found vendor-led partnerships were one-third more likely to succeed, showing targeted workflow tools outperform broad custom builds.
Meanwhile, others look at the report claiming 95% of enterprise GenAI investments yield zero returns with skepticism. It cites $30–40 billion in spend and surveys of 52 firms and 153 leaders but offers no supporting data, defines “success” narrowly, and concedes figures are only “directionally accurate.” Lead authorship from a Microsoft scientist raises questions about MIT’s role, and critics say the core claim lacks evidence.
Stanford researchers Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen published a report finding that early-career workers (ages 22–25) in the most AI-exposed jobs have seen a 13 percent relative drop in employment, while older and less exposed workers remained stable or grew. Losses came through job cuts, not pay, suggesting generative AI is disproportionately impacting entry-level roles.
ByteDance is launching a share buyback valuing it at over $330 billion, up from $315 billion six months ago. Q2 revenue reached $48 billion, topping Meta, though TikTok’s U.S. arm remains unprofitable and faces a September divestment deadline. Unlike peers, ByteDance funds buybacks from its own balance sheet while investing heavily in AI infrastructure.
In a new essay, AI researcher Shunyu Yao argues that AI’s “second half” shifts from new training methods to defining and evaluating tasks. With language pre-training, scale, and reasoning, RL now generalizes across domains, making benchmarks quickly obsolete. The frontier is real-world utility; new evaluation setups, human-in-the-loop interaction, and long-term tasks over incremental algorithms.
At Contrary Research, we’ve built the best starting place to understand private tech companies. If you're interested in researching and writing about tech, apply here for our Research Fellowship!