Dharmesh Shah says software survives, but the public grid pays

By Imani Sutton · Reporting from Atlanta ·

On a16z, Dharmesh Shah argued that software startups will thrive by selling artificial intelligence, but turning enterprise tools into metered digital utilities leaves everyday ratepayers funding Silicon Valley's massive power demands.

Dharmesh Shah has three rules for business: never own a restaurant, never compete with Sam Altman, and never be a landlord and charge rent. Speaking with host Joe Schmidt on the a16z podcast, he argued why standard business software will survive the rise of artificial intelligence. The HubSpot co-founder and chief technology officer delivered that second rule with an easy chuckle. Listening to Shah describe the future of his industry, the irony is hard to miss. Silicon Valley is actively preparing to turn every piece of software into an unregulated digital landlord.

On the podcast, Shah argued that software companies face an unprecedented opportunity rather than extinction. Drawing on his experience building HubSpot over two decades, Shah dismissed fears that frontier artificial intelligence labs like OpenAI or Anthropic will swallow every business tool. Instead, Shah contended that basic intelligence models are rapidly commodifying into interchangeable infrastructure. In Shah's view, the durable value belongs to application software companies that package those models into everyday workflows, proprietary customer context, and what he calls a "second brain." To prove it, Shah pointed to his latest project, Uspot, a CRM for solopreneurs built inside HubSpot's internal incubator. He noted that more than 70 percent of new American businesses tracked by census data are expected to remain single-person operations eager to swipe a credit card for metered artificial intelligence.

The illusion of the cheap utility

The strongest case for Shah and Schmidt’s optimism rests on the mechanics of software distribution. To their credit, they correctly identify the fatal flaw of many artificial intelligence startups: trying to outrun foundation labs down the center of the road. Foundation models require billions in computing capital, but as Schmidt observed, application layers that control customer relationships hold durable ground. If raw model output becomes a fungible commodity routed through intermediaries to get the lowest price, application companies avoid platform lock-in. They hand small businesses powerful automation that once required entire departments.

Yet that entire optimistic case collapses the moment you examine what this software actually consumes. Shah casually floated the idea of spending $200,000 on computing tokens just to have an automated agent rewrite an entire codebase in an alternative language to test architectural choices. But computing tokens are not magic dust. They represent physical compute cycles running on server racks inside industrial data centers owned by conglomerates like Microsoft, Google, and Amazon. When Shah insists that artificial intelligence is "not some new energy source" but "just software," he glosses over the immense physical appetite of the machine. The hundreds of billions of dollars spent on pre-training and server farms require real power from regional electric grids. They also require millions of gallons of evaporative cooling water from local watersheds.

The long record of platform rent

History shows exactly what happens when tech evangelists promise that a new layer of infrastructure will decentralize power. Corporate computing moved from desktop mainframes to the web, and later from private servers to Amazon Web Services. Silicon Valley promised that cheap hosting would democratize enterprise forever. Instead, capital reconcentrated into a handful of cloud landlords who extracted permanent rent from every business operating online. For this era to play out differently, the frontier labs spending hundreds of billions on physical infrastructure would have to quietly accept low margins forever. Platform history shows they never do. Once workflows are locked into automated agents, the price per token climbs.

The bill for this frictionless future will not be paid by venture capitalists or software executives running late-night experiments. It will be paid by self-employed workers watching their monthly software subscriptions convert into variable consumption meters. It will be paid by working families whose residential electric bills rise to subsidize the grid upgrades required by corporate data centers. Citizens heading to municipal council meetings and public utility commission hearings have direct leverage here. When local boards consider tax abatements, zoning permits, and utility rate cases for massive server facilities, voters can demand that tech companies finance their own dedicated renewable power, pay industrial tariffs that shield residential ratepayers, and stop shifting the physical cost of computing onto the public grid.

Sources

  1. Why HubSpot’s Co-Founder Is Betting Against the SaaSpocalypse