Topics Pioneers Insight Method Author
Back

The U.S. Votes on Data Centers, Big Tech Impact, Google’s AI Response | Sharp Tech with Ben Thompson

·Sharp Tech

Rendering the long image…

Topics AI Data Center Power Bottlenecks Trillion-Dollar AI Capex and Bubble Risks Commercial Shifts in Mobile Recommendation Algorithms One-Person Companies via End-to-End AI

Summary

  • The data center backlash has reached the industry’s presumed safety valve, Texas. Sharp reports that 43 of 44 local measures opposing construction failed, while recalls warned officials that approving projects could cost them their jobs. Texas had been expected to house 25% of US capacity by 2030—the answer to opposition elsewhere was “we’ll just build it in Texas”—but its incoming governor campaigned on a moratorium. Existing construction cushions 2027–2028; the greater concern is 2029 onward.
  • Tech’s political problem is distrust of the future it is selling, which fact checking alone cannot resolve. Thompson argues that “what you’re actually dealing with is sentiment,” reinforced by messaging that AI will replace workers and concentrate wealth. Separately, Sharp urges companies to compensate communities—“If you want to get a data center built, write checks. Put money in people’s pockets”—but asks whether that can overcome philosophical opposition.
  • Restrictions could entrench the hyperscalers that opponents want to punish. Existing sites, power connections, cash generation, cheaper credit, and stronger equity currencies give incumbents advantages over frontier labs seeking new infrastructure. Thompson’s illustration: $200 billion of Amazon capex against $180 billion of cash generation still leaves an enormous funding advantage over loss-making labs. OpenAI and Anthropic might endure while ownership changes: “The question is who are the equity holders?”
  • An externally imposed capex slowdown could give revenue time to catch up with investment. Thompson’s conditional upside is that easing capital pressures could let “everyone” win. Nvidia could sell fewer chips in absolute terms while strengthening its relative position: constrained power makes “the most tokens per watt” worth a premium. Existing neocloud capacity could also appreciate even as expansion becomes harder.
  • Google’s search disappointment could represent a monetization transition rather than a terminal business failure. Alphabet’s reported decline of more than 13%, erasing over $500 billion, followed disappointing search and advertising results. Thompson sees a possible replay of Meta’s Stories and Reels transitions: attention moves before advertising tools and monetization catch up. Sharp’s objection is substantial—Google must replace “the greatest business that has ever existed”—and Thompson acknowledges his poor forecasting record on Google.
  • Google may be using search to finance an infrastructure transformation. Sharp compares search’s potential role with See’s Candies funding Berkshire Hathaway’s railroads: lower margins can accompany greater absolute profits. Thompson’s “Google Capital Company” idea is that Google could become an AI infrastructure provider while transforming search into an interface and eventual agent. Whether that ultimately makes financial sense remains uncertain.
  • Reddit’s roughly 30% selloff illustrates the cost of losing Google’s distribution bargain. Sharp regards management’s direct-destination argument as optimistic spin; Thompson calls it a “valiant attempt to state reality.” The response is higher revenue per user through logged-in experiences and better advertising, alongside licensing continually refreshed content. The familiar 1% posting, 9% responding, 90% lurking split matters: contributors already log in.
  • AI’s more credible entrepreneurial payoff may be many $100,000 businesses rather than a few billion-dollar solo companies. A school-lunch app illustrates software tailored to a small community rather than a mass market. Thompson advises using AI as “a machine gun you can train on your weaknesses,” freeing time for strengths. Payments, authentication, hosting, and security remain unresolved opportunities around this “group chat” era of software.

Deep dive

1. Texas’s role as the automatic escape route is now uncertain

  • Sharp reports that 43 of 44 local measures opposing data center construction failed; North Dakota was the exception. Thompson considers recalls of mayors and commissioners particularly consequential: officials now know approvals can cost them their jobs, including decisions that never reach a public ballot. In the Madison market, candidates from both parties attacked each other over alleged data center support.

  • Texas is the larger shock. Gina Hinojosa was elected governor after promising a moratorium on new data center construction until the legislature could act; Thompson notes that the legislature meets only once every two years. The episode cites a projection that Texas would hold 25% of US data center capacity by 2030, with expectations of a still larger share over time.

  • Thompson recalls a senior industry participant dismissing a canceled Wisconsin project because everything could move to Texas. Abundant natural gas, including gas otherwise flared, had already supported Bitcoin mining; Texas also operates its own grid alongside the eastern and western grids. Land, energy, and a presumed friendly government made “we’ll just build it in Texas” the backstop. That political assumption is now unsettled.

2. The backlash targets AI’s promised future

  • Thompson’s framing: misinformation often supplies ammunition for an existing preference rather than creating it. “You’re trying to refute facts when what you’re actually dealing with is sentiment.” Debunk one water or electricity claim and another objection can replace it, because the underlying force is distrust of technology and discomfort with AI. Thompson sees people voting against “this version of the future.”

  • Sharp offers the slogan, “If you love advertising, you’re optimistic about AI.” Thompson explains the connection: good advertising identifies a problem people have learned to tolerate and shows a solution. Consumer AI companies have failed to make that connection convincingly. Sharp’s pushback is that they are building exceptionally expensive technology while their clearest public application remains replacing large portions of the workforce.

  • The deeper mistake, Thompson says, is treating people as interchangeable inputs ranked on an IQ scale. Humans are “need generation, desire generation machines”: they invent wants, businesses arise to satisfy them, and demand changes. Job-replacement messaging implicitly tells the public, “All of you are worthless.” Sharp finds the backlash understandable given that message and the concentration of wealth among big tech companies.

  • Sharp leaves durability open: movements with a memetic character can fade over two or three years, although those years matter enormously for construction. Thompson thinks broader anti-tech sentiment will persist and data centers will remain on ballots. The globalization comparison sharpens his point: people could not vote when factories closed; here, they have a veto over a future they dislike.

3. Communities need benefits before the virtuous cycle arrives

  • Thompson calls corporate America’s underlying error “the reduction of morality to legality”: exploiting every permissible loophole invites regulations that constrain innovation alongside abuse. Amazon’s headquarters auction exemplified the posture—one of the world’s most powerful companies soliciting subsidies from competing cities. Data center companies should have anticipated public unease and demonstrated community benefits before demanding inducements as the price of locating there.

  • Loudoun County, Virginia, supplies the affirmative example. Sharp identifies it as the country’s richest county; Thompson says its property taxes have gone down. Sharp adds, with an explicit hedge, “I believe power prices have gone down.” Sharp traces the county’s growth to historic connectivity, proximity to Washington and the populous East Coast, access toward Europe, and available power. Thompson adds that the resulting industry developed a specialized construction ecosystem.

  • Thompson emphasizes that Loudoun represents a roughly 30-year virtuous cycle. Sharp cautions, “You can’t sell the long run” while voters bear immediate costs; Thompson adds that communities should be paid for the privilege of running up their electricity bills. Sharp distinguishes election incentives from governing incentives: officials also risk watching valuable industrial projects go elsewhere.

  • Sharp predicts communities will reverse the negotiation: “You want a transactional model, you’re going to get a transactional model good and hard.” His practical prescription is direct compensation—“If you want to get a data center built, write checks. Put money in people’s pockets.” He agrees companies can accept that baseline but doubts it guarantees approval where residents oppose AI philosophically or resent tech’s wider effects on society.

4. Existing projects buy time, while new investment may move abroad

  • The hosts distinguish new permitting from projects already underway, citing roughly 130 Texas projects under construction. Thompson expects much of that capacity still to arrive in 2027 and 2028; the concern shifts toward 2029 onward. That limits the immediate construction impact without removing the possibility that today’s political decisions create a later bottleneck.

  • Going abroad is “probably likely to happen,” Thompson says; the hosts cite a $15 billion Google project in Malaysia. Canada and Mexico enter the discussion, but permission, power, financing, and security still matter. Thompson had expected substantial Middle Eastern construction. He recalls an AWS region in Bahrain, he thinks, whose customers were told their data was gone forever—an example raised amid concerns about security.

  • The possible domestic dividend is power generation that survives an AI investment bust. Sharp says the US has been living in “power poverty” for roughly 50 years, dating back to the oil crisis, and contrasts today’s efficiency mindset with an earlier assumption of abundant energy. If economically extravagant AI spending produces enduring generation capacity, its long-run value could be enormous. That payoff remains conditional, but Sharp believes substantial power projects are already moving toward completion.

5. Space infrastructure becomes less absurd when launch capacity already exists

  • Thompson rejects the image of warehouses floating in orbit. His proposed unit is an individual rack: an NVL72 is approximately the size of a current Starlink satellite, he says. Cooling remains difficult, requiring massive radiators, but satellite deployment and laser communications already exist. Agent-serving inference could also tolerate more latency than a person waiting for an immediate chatbot response.

  • The economic starting point matters. SpaceX is already developing Starship and expanding launch capacity; Sharp argues that this creates a need for useful payloads. Sharp characterizes Starlink and SpaceX as “the ultimate hammer looking for a nail company.” He argues that launch can be treated as a marginal cost for some decisions, while explicitly acknowledging that it is not actually free. Starlink’s apparent profitability also depends partly on internal launch pricing rather than an arm’s-length premium.

  • Thompson’s shipping analogy separates operating economics from sunk investment: a container costing $1,000 at the margin can rationally ship for $1,100 even if depreciation makes its accounting cost $2,000. Reported losses can coexist with positive cash flow; debt service can ultimately cause bankruptcy. COVID’s $17,000 containers illustrate why commodity operators endure lean years—and why investing only after prices recover can be too late.

  • Sharp argues that launches build on the learning function; Thompson adds that Musk’s overarching goal is Mars. Verizon, AT&T, and T-Mobile should worry about a rival whose seemingly irrational cellular economics support a different objective, Thompson argues. Sharp frames orbital computing as a “5 to 10 to 15 year solution.” Thompson emphasizes that most data centers will still be built on Earth for the foreseeable future.

6. Hyperscalers hold the strongest cards in a constrained buildout

  • Frontier labs have two potential integration paths, Thompson explains: combine models with consumer-facing applications, or combine models with purpose-built chips and infrastructure. The latter requires capital and new facilities. Anthropic’s Broadcom chip agreement does not remove the need to construct places to deploy those chips. Existing hyperscalers begin with assets and relationships that the labs must assemble.

  • Legacy facilities are not easily converted into modern AI facilities: older designs generally use air cooling and lower density, while current GPUs require much denser, water-cooled installations. Conversion can amount to rebuilding. Nevertheless, incumbents already control land, power connections, local relationships, and sites communities accept. Sharp cites Meta discussing expansion of existing facilities as an alternative to entirely new construction.

  • Thompson’s financing illustration is Amazon spending roughly $200 billion on capex against $180 billion of underlying cash generation. Exceeding that cash generation does not erase the advantage of having it. Frontier labs start from negative cash flow. Successful existing businesses also support cheaper borrowing, allowing Amazon to obtain attractive Anthropic arrangements partly because it can finance infrastructure more cheaply than its customer.

  • Equity is the final funding source, and incumbents have an advantage there too. Buying newly issued Amazon shares purchases exposure to a successful existing business alongside speculative AI investment; buying Anthropic equity depends on Anthropic succeeding. Hyperscalers also possess greater lobbying capacity and more resources to compensate communities. If those become decisive construction inputs, opposition reinforces incumbent advantages rather than weakening them.

7. A compulsory capex pause could relieve the financing race

  • Nobody can easily slow down voluntarily. Amazon risks ceding capacity to Google; Microsoft’s brief restraint already produced shortages, Thompson argues. Frontier labs face the same competition in training: exceptionally cheap inference does not eliminate the need to spend heavily to remain at the frontier. Anthropic cannot independently decide that OpenAI, Google, or another competitor will stop advancing.

  • An external constraint could let everyone pause without surrendering competitive position. Canceled projects could mean flatter capex, relieved shareholders, higher share prices, and a more valuable equity currency for subsequent fundraising. “If the revenue comes,” Thompson stresses, those financing pressures can unwind positively. A slowdown that gives revenue time to arrive could therefore let “everyone” win.

  • Sharp asks whether compute-starved labs eventually approach their suppliers hat in hand. Thompson does not expect OpenAI or Anthropic to disappear: “The question is who are the equity holders?” Google and Amazon owning more of Anthropic is one possible outcome. He also warns against treating contracted demand as unconditional protection: a guaranteed buyer that is itself OpenAI can transmit financing trouble through the chain.

8. Scarcity changes Nvidia’s economics and complicates the China comparison

  • Nvidia faces a straightforward short-term problem: fewer places to install chips can mean fewer sales. Relatively, however, power and facility constraints favor the supplier delivering “the most tokens per watt.” Paying its premium becomes easier to justify when scarce infrastructure must produce the greatest possible output. Thompson distinguishes being “better absolutely or better relatively”; a prolonged scarcity regime can hurt volume while strengthening competitive positioning.

  • Existing neocloud facilities present the same tension. Slower additions make installed capacity more valuable, but restrictions also make expansion harder and sufficient scale less attainable. Analysis that assumes intelligence is already a commodity skips a necessary condition: effectively unlimited compute. Thompson thinks the scarcity period has been extended, making conclusions drawn from an abundant-compute end state less useful for current decisions.

  • China is a conditional longer-run concern. America has a “huge chip advantage,” Thompson says, but China has been pushed to solve chips domestically; if it succeeds, greater power and construction capacity could eventually produce a manufacturing-like scale advantage. Sharp’s disagreement matters: despite acknowledging those possibilities, he has “never been less concerned” about US-China competition and is “not that panicked” by claims voters handed China the future.

9. Google’s search miss resembles Meta’s monetization transitions

  • Sharp cites Bloomberg’s report that Alphabet fell more than 13%, losing over $500 billion in its largest market plunge. Overall sales missed projections, as did search and related advertising, a business representing roughly half of revenue and financing AI investment. Search advertising growth slowed to its weakest pace in five quarters. Thompson’s response is explicitly “a very optimistic read” of the earnings call.

  • Stories and then Reels moved users away from Meta’s already monetized surfaces. Advertising tools, measurement, and advertiser habits lagged behind attention, creating bargains for early advertisers. Meta eventually built the systems to monetize the new inventory. Thompson’s rule: “You want to buy a digital advertising company when its inventory is exploding.” Reels produced the spectacular “blip” associated with Facebook shares around $83.

  • Google’s language sounds familiar: get the product right, distribute it widely, then monetize. AI Mode can increase queries and engagement while diverting users from the search-results pages where existing ads appear. More time on Google therefore does not automatically mean more revenue. The bullish interpretation is that product adoption has advanced faster than the advertising format, rather than that demand has disappeared.

  • Sharp’s pushback: search was “the greatest business that has ever existed,” a much harder benchmark than merely finding profitable chatbot ads. Thompson grants the question but sees “a decent chance” that monetization makes this another temporary blip. Sharp notes Thompson’s strong Meta record and weaker Google record; Thompson jokes, “You should always do the opposite of what I say about Google.”

10. Google may be converting search cash into a different company

  • Thompson’s “Google Capital Company” framework treats search as funding for a future infrastructure business: data centers, proprietary chips, cloud, and stakes in companies including SpaceX and Anthropic. Sharp compares the idea with See’s Candies financing Berkshire Hathaway’s move into railroads. Railroads have worse percentage margins but can generate much larger absolute profits. Search could prove to have been “just the fuel to get the real business.”

  • Google is also making a separate bet on transforming search itself. “Who’s done more to obliterate the search experience than Google itself?” Sharp asks. Rather than protecting the old interface, Google keeps moving users into chatbot interactions. Thompson’s prediction is straightforward: “Search is going to be your agent,” eventually doing tasks such as booking things rather than merely returning links.

  • Thompson admires the execution and willingness to choose, despite uncertainty over ultimate financial returns. Sharp says Google’s cloud numbers are incredible and its profit margins are skyrocketing; Thompson adds that infrastructure barriers to entry are enormous. Sharp notes Thomas Kurian’s appearance on the earnings call and reads Google’s choices as a sign of priority. Thompson says that without the “Google Capital Company” trajectory, he would be significantly more worried about Google.

  • The erosion is visible in personal usage: Sharp now uses ChatGPT for 80–90% of tasks previously handled by Google, and cites OpenAI saying search-like ChatGPT queries equal one-fifth of Google’s volume. They consider whether Google has been extracting extra revenue from a declining franchise. Thompson calls that plausible but explicitly does not know how accurate the cited critique of its tactics is.

11. Reddit must replace referral volume with value per user

  • Sharp reports Reddit losing roughly 30% of its share value over about a week. Steve Huffman’s explanation is that people still search for Reddit, but Google sends them there less often: “Google is teaching people that if you want to come to Reddit, you need to come to Reddit.” Sharp calls the direct-destination argument valiant spin; Thompson calls it a “valiant attempt to state reality.”

  • The wider bargain is breaking down. Google indexed publishers to improve search, sent them traffic, and provided AdSense to monetize it. Thompson distinguishes possible abusive practices from the fundamental business problem: “Publishers’ problems are not Google’s ads.” Moving users into AI Mode removes distribution that supported the broader web economy. The end is of the internet “as we know it,” organized around one large market and scale.

  • Reddit consequently needs to “maximize your revenue per user” rather than assume unlimited low-value referral traffic. Thompson does not predict a paid model; he expects better ad monetization through logged-in users, customer understanding, and targeting. This requires changing the measures of success as well as the product. The business is being reoriented by external reality, whether management welcomes that process or not.

  • Its durable asset remains people producing fresh material for free, which also supports valuable LLM data licensing. Thompson invokes the 1% who post, 9% who respond, and 90% who lurk: the contributors already log in. Losing some anonymous visitors need not destroy the content engine. Requiring accounts might even encourage participation, although he acknowledges that the transition will be painful.

12. Small software gains usefulness without becoming a mass-market business

  • A listener’s school-lunch app captures the “group chat era” of software: public information is hard to find, while an app with notifications fits parents’ existing habits. A school district’s own app might accumulate unrelated functions. The opportunity is a simple tool tuned to a particular community—not necessarily software that millions will buy when millions of others can build similar things.

  • The listener asks who supplies payments, authentication, hosting, and security: platform owners, AI labs, Vercel and Supabase, or new businesses. Thompson offers no settled winner. Sharp highlights Google Play’s recently introduced friends-and-family app distribution and contrasts it with TestFlight’s developer-account and enrollment friction. Sharp sees Google ahead on this particular issue; Thompson agrees that this is what’s needed. Current improvised hosting and security arrangements expose substantial unsolved opportunities.

  • Small scale also restores motives beyond profit, Thompson argues. Making a useful neighborhood app can earn gratitude and status. Duplicating similar apps across many communities looks inefficient, but customization can improve welfare. His analogy is fashion: people are “very inefficient about wearing clothes,” yet their changing desires support an industry. Human-generated differences can create value rather than merely represent costs to eliminate.

13. AI’s practical upside starts with weaknesses and modest businesses

  • Thompson’s advice to MBA students is deliberately sharp: “If you see AI as an opportunity to do your job, you’re in trouble.” Instead, protect the strengths people pay for and use AI on secondary work. “AI is like a machine gun you can train on your weaknesses.” The purpose is to redirect time and attention toward what someone does particularly well.

  • Thompson is unsure about predictions of billion-dollar one-person companies, while allowing they may happen. What he “definitely” believes is that many more people can run $100,000 businesses. Sharp raises the example of a focused local-news publication; Thompson says its cost structure could fit its audience, unlike a newspaper’s inherited overhead. Removing costs lets modest revenue support a livelihood and makes more experiments economically possible.

  • Platforms may follow demonstrated demand rather than precede it. People could first assemble their own small businesses and software stacks; a Substack-like provider could then make the process accessible to many more. Sharp says he had considered building a Substack but doubted how many writers could earn money. Thompson’s response is that people underestimated the opportunity: “Turns out a lot.” His optimism rests on people’s underestimated ability to create markets when the required revenue falls.