The Thread: AI’s Growing Pains Are Real and Expensive
The stories from the past two days share a single, quiet thread: the material cost of artificial intelligence is rising just as quickly as its capabilities. From Google’s new, harder-working model with unchanged pricing to Amazon’s new anti-scam feature that leverages AI, and from Fervo’s 400-megawatt geothermal deal with Google to Broadcom’s strong but cautious earnings, one pattern emerges. AI’s next phase is not purely a software story - it is a story about power, chips, deployment complexity, and consumer trust. And for US companies and consumers, these costs will be felt in prices, reliability, and security.
Powering the Boom: Geothermal for Data Centers
The most direct signal comes from the energy side. As TechCrunch reported, Fervo has signed a deal to supply Google with 400 megawatts of enhanced geothermal power, with an option to expand to 1 gigawatt - enough to run a very large AI data center in Utah. That is not a minor footnote; it is a marker of how much electricity modern AI workloads demand. Data centers are the physical engine room of AI, and the US grid is not growing fast enough to power them on conventional sources alone.
The fact that Google is turning to enhanced geothermal - a relatively young technology - indicates that AI’s energy appetite is changing procurement strategies at the largest US tech firms. It also means that the cost of electricity, and the infrastructure to deliver it, will become an input cost for AI services. For US consumers, this is not an abstract concern. AI-powered products - from search to assistants to cloud services - depend on cheap, reliable power. If that power is harder to source, the price of those services will eventually rise, or the services will be limited to regions where power is abundant. The Fervo deal also points to a broader trend: tech companies are becoming direct investors in energy infrastructure, not just customers of utilities.
Chips: The Cost of Customization
The second cost center is silicon. Broadcom beat expectations on earnings and revenue in its third-quarter report, as SiliconANGLE noted, and hinted at more business with leading AI labs. The chipmaker’s stock dipped after its current-quarter forecast trailed Wall Street’s estimates. The key word here is "custom chips." AI labs are increasingly moving away from off-the-shelf GPUs to custom accelerators designed for specific workloads. That customization is expensive - both to design and to manufacture.
Broadcom’s strong third quarter but cautious fourth-quarter outlook suggests that the AI chip market is not a straight line upward. There is a limit to how fast even the most AI-hungry companies can absorb new hardware. The short-term dip in Broadcom’s stock after a beat is a reminder that Wall Street expects growth, but the underlying costs of custom chip development are rising. For US technology companies, this means that AI hardware will not get cheaper on a predictable curve. The days of exponential performance gains per dollar may be flattening. For US consumers, the cost of AI will be embedded in the price of everything from smartphones to cloud subscriptions, because the chips inside them are becoming more specialized and less commodity-like.
Models: More Work, Same Price - For Now
Google’s new Gemini 3.8 Flash model, as The Verge reported, "works harder" than its predecessor by performing more reasoning steps and calling tools iteratively. It launched just weeks after 3.7 Flash and retains the same introductory pricing - $0.75 per million input tokens and $3.75 per million output tokens. That is a notable pricing decision. The model does more, but the per-token cost is unchanged. That suggests Google is absorbing the additional compute cost in the short term, likely to gain market share in developer tools.
But that pricing is introductory, and the word "introductory" carries weight. If the model genuinely requires more compute per request, that will show up in Google’s infrastructure costs. Either the company will eventually raise prices, or it will find efficiencies in inference and hardware. The "works harder" language is also a signal to developers: you can get better performance, but it may not come free forever. For US businesses building on Gemini, this is a cautionary note. The current price is a teaser, not a floor. For consumers, the cost is indirect but real - more capable AI assistants and search features will raise the cost of running Google’s services, which could translate into higher subscription fees or more advertising in free products.
