After a research trip to San Francisco, Fidelity technology analyst Jonathan Tseng  (bottom left) and US equities portfolio manager Sam Thomas discussed the changing economics of artificial intelligence in a late-July 2026 webcast titled ‘What Is Silicon Valley Saying About Cost (and Jobs)?”

They noted that the conversation surrounding artificial intelligence has shifted from theoretical wonder to a gritty, high-stakes debate over the bottom line. It is no longer just about what the models can do, but also about what they cost to run and who they might eventually replace.

For investors, the central question is whether AI-driven productivity and revenue can grow fast enough to justify the industry’s rapidly rising capital expenditure.

The price of progress and the $800bn bet

While headlines scream about the ruinous expense of AI, the reality on the ground is more nuanced. Tseng said token costs at major internet companies often remain in the low single digits as a percentage of total expenditure, leaving companies considerable room to increase their AI usage before costs become problematic. “However, at 20%+, the CFO has a problem,” he said.

The scale of infrastructure spending is staggering. Just one year ago, estimates for aggregate capital expenditure sat at $400bn; today, that figure has doubled to $800bn, according to Thomas. For hyperscalers like Google, Microsoft, and Amazon, the “hamster wheel” of spending is turning faster than ever, with aggregate capex now essentially equal to aggregate free cash flow. “[The difference] got to zero much faster than expected.”

To manage these costs, companies are resorting to “valuemaxxing,” choosing smaller, efficient models for mundane tasks like email summaries rather than wasting “Einstein-level intelligence” on basic administrative work. Finding the right model for a specific task is a genuine challenge for companies. Thomas suggested that open-source models could offer executives a more economical solution.

Redefining the office and the end of “body shopping”

Fears of mass unemployment are often met with historical perspective by industry experts. Tseng framed AI not as a ‘killer robot’ but as the latest stage in the evolution of computing—a progression from abacuses to mainframes and, now, “fuzzy logic.”

The shift is moving from automating manual labour to “automating ideas” and creative labour. This transition spells trouble for “body shopping”—the supply of undifferentiated human labour—which is now seen as low-hanging fruit for smart models.

Yet companies such as Sierra illustrate how AI can also create demand for human ingenuity—for example, in designing and deploying customer-service agents that improve productivity. The goal is to do “more with more” or “more with the same,” rather than simply cutting staff .

Robots, demographics and the physical frontier

The next frontier of this revolution is not confined to a screen; it is moving into the physical world. Analyst Kitty Yang (top right) argues that automation is becoming a “need and not a choice” due to shrinking working-age populations in Japan, China, and the US

Advances in sensors and motion control have brought industrial robots to an inflection point, enabling them to perform far more complex tasks than before. Consequently, Yang noted that

AI is moving beyond simple repetitive tasks to handling complex industrial and medical applications using “digital "twins"—virtual representations of physical systems that allow companies to model and test changes before implementing them in the real world.

Tseng highlighted the emergence of physical products, such as Samsara’s fleet-management systems and Axon’s police cameras, that use AI to automate paperwork and improve safety. For investors, this creates “embedded option value” in suppliers of the connectors, sensors, and radar systems that allow robots to “see and feel” the world around them, said Thomas.

Geopolitics and the quest for AI sovereignty

As AI becomes a tool of national power, “AI sovereignty” has entered the lexicon. Recent US government clampdowns on exporting the most powerful models have raised fears of a “self-inflicted wound.”  Thomas argued that restrictions on American models could inadvertently provide a “shot in the arm” for sovereign champions like Mistral in Europe or force global players toward Chinese open-source alternatives.

“The weaker your model is, the more you want it to be open-source,” stressed Tseng. As developers build on other people’s ideas, successful models tend to become more closed.

Economist Edoardo Cilla (bottom right) suggested AI could lift long-term GDP growth by 40 to 50 basis points, though this relies on massive upgrades to energy grids. Countries seeking to lead in AI will need abundant, reliable and competitively priced electricity, as well as substantial investment in their power grids.

AI capex begins to pay off

Despite these constraints, investment and adoption continue to accelerate. Revenue growth in the cloud sector is accelerating, suggesting that last year’s capex is already bearing fruit.

Thomas cited Anthropic’s rapid revenue growth as one source of demand for hyperscaler capacity and advanced chips. Although this demand has not yet flowed fully into hyperscaler revenue, he said activity at the chip level indicates that AI developers and infrastructure providers are expanding faster than expected six months ago.

The outlook for AI adoption

AI adoption is spreading beyond early adopters and software developers into the broader white-collar workforce and management ranks. While the transition may be disruptive, the long-term outlook remains optimistic. AI could eventually generate new occupations and sources of demand, but the timing and distribution of those gains remain uncertain.

Tseng, nevertheless, remained optimistic about the longer term: five years from now, after the most disruptive phase of the transition, “the future will be bright.”