The real AI pinch point
Hint: it’s not chips
This week, we’re talking:
$109B went into AI last year. The real choke point now is power… literally.
Compliance is splintering. That’s gonna be more than a headache for builders.
Jane Goodall died this week. Lots of ink has been spilled on her legacy, not enough on what she really was: a disruptor.
A new study out of Illinois shows how conspiracy rabbit holes end marriages.
YouTube settled with Trump for $22M. One GIF sums up my feelings.
Ezra Klein and Ta-Nehisi Coates clashed in print over that Charlie Kirk op-ed. When they met on the mic, the fight turned into a reckoning.
My Take:
U.S. VCs put 109 billion dollars into private AI in 2024. That is about 12 times China and 24 times the UK. The money is loud, but the real deciders now are energy, regulators, and geopolitics. I went rapid-fire with my friend, George Lee, Co-Head of the Goldman Sachs Global Institute, about how this actually plays out for builders.
Tom Chavez: The U.S. outspends China 12 to 1. The Middle East is suddenly in every AI conversation. Mirage or movement?
George Lee: On China, you gotta be careful with the headline ratio. Transparency is limited, and a lot of spend sits in adjacent buckets like robotics, drones, data, and energy infrastructure. They were ambivalent about generative AI at first, but they have pivoted. After DeepSeek, the signal is clear. The irony is a relatively closed system leaning into open source and open weights to project soft power across the Global South. On the Middle East, do not underestimate them. Abundant capital. Abundant energy. The ability to coordinate fast. Top down can be a bug and a feature when the cycle is moving this quickly.
Tom: But do the rulers let the market breathe, or is this planned economy cosplay?
George: Today it is more top down. In a fast, resource intensive wave, that can be an advantage. Whether it becomes a constraint later, we will see. This generation of rulers is ambitious and progressive about AI. They see it as the motive power of the next industrial revolution.
Tom: You use a canal and locks metaphor for the AI buildout. Not one bottleneck, a sequence. Where are we now?
George: Early it was silicon. Then advanced data center capacity. Now the pinch point is energy. AI class data centers are massive power consumers. That is the lock we are in.
Tom: Climate math is not forgiving. What is the practical fix now, not five years from now?
George: Two pieces of context. First, U.S. electricity demand has been basically flat for about 20 years. Efficiency gains and a move from dispatchable to more intermittent sources did a lot of work. Now we have to scale supply and the know how to run that playbook at speed. Second, you can already see stress. One U.S. grid printed roughly 22 percent year over year wholesale price increases. Once ratepayers feel that, public utility commissions get involved. That is a policy brake, not just a technical one. Near term, think flex power. The grid is built for peak. Off peak there is slack. If AI workloads flex, curtail during peaks, and shift training and ETL to slack periods, you can harvest serious capacity as a temporal bridge while new generation catches up. A Duke analysis pegs the latent pool at roughly 75 to 125 gigawatts if we line up incentives and orchestration.
Tom: Not just theory, right?
George: Right. Google just signed agreements with utilities to curtail or shift AI load at peak in exchange for better access and pricing off peak. Expect time of day deals to become standard. And if the U.S. cannot meet the moment, expect spillover. The Middle East and Canada, plus other power abundant or faster permitting regions, will host the next wave of advanced AI data centers.
Tom: Translation for users: sometimes I might wait an extra 200 milliseconds.
George: Exactly. Users already show higher latency tolerance in chat. Old search trained us to bail if results took longer than about 0.6 seconds. Chat flipped that script. We accept a thinking beat. Training and reinforcement runs are easy to snapshot and shift. Keep inference responsive. Flex almost everything else.
Tom: I feel this in my own workflow. Fifteen years ago Google bragged about 0.27 seconds. We all got spoiled. Now I usually have three chatbot tabs open. One cranking code, one on a deep research query, one drafting. I let them cook 10 to 12 minutes. It is plate spinning at the carnival. Start one, move to the next, circle back when the first one is cooked.
George: Exactly. Humans are adapting to machine rhythm, not just the other way around.
Tom: Enterprise reality: Non determinism freaks people out. Same prompt, different answer tomorrow. Deal breaker?
George: It is coexistence. Deterministic systems will keep running transactions. Repeatable, traceable, correct. Probabilistic systems will supercharge the big area under the curve: summaries, copilots, research, ops assist. The maturity model shifts from one and done UAT to continuous sampling, testing, and tuning. Think continuous manufacturing, not discrete. Also, these models are the most human machines we have built. They inherit our foibles. Ask you the same question three days in a row, you might give three different answers. Use them where that is acceptable, and keep deterministic rails for the rest. Finance has lived with probabilistic models for decades. That muscle memory helps.
Tom: I tell friends to ask the same question again tomorrow. When the answer differs, that is not the system breaking. That is the paradigm. Build guards and workflows around it.
George: Exactly. Different answers are the point. Use guardrails and continuous QA so the variability works for you, not against you.
Tom: Regulation: Helpful guardrails, or a culture war tripwire?
George: Both are possible. Risk one is a headline driven overreaction to a bad event that produces fast, blunt rules. Risk two is politicized pressure on model outputs. These systems are cultural transmitters, so governments care a lot. The White House and others are already focused on that. The whole woke AI fight is a preview. Too heavy a hand, or none, both create problems. The encouraging bit is that more regulators are learning the language of probabilistic computing and engaging constructively.
Tom: Say I want to ship worldwide. What actually trips me up?
George: Fragmentation. I once thought we were converging on harmonic standards. Not anymore. Data locality, sovereignty, and linguistic and cultural constraints are intensifying. Even giants are opting out of markets they cannot or will not comply with. Builders have to pick geographies, design for modular compliance, and accept that global on day one is mythology.
Tom: M and A: What is actually happening behind the headlines?
George: This platform shift is unusual because resource intensity and concentrated talent currently favor incumbents. Usually platform shifts hurt incumbents. This one, for now, helps them. The most visible motion is not classic roll ups yet. It is acqui hires and superstar researcher contracts at prices that used to be whole company numbers. If you are spending 100 billion dollars in capex and a small team can boost utilization a couple of points, the math can pencil. Expect strategy separation among model leaders. For example, Anthropic leaning into coding and engineering, OpenAI leaning more consumer. That divergence will inform M and A.
Tom: A lot of what you just described sounds like pro sports economics sneaking into tech. Am I reading that right?
George: That is a fair read. Superstar packages and acqui hires are filling the gap before classic consolidation.
Tom: That sportsification of engineering makes me… twitchy. Forty people in the magic room win the lottery. The other two hundred are left holding the bag. That breaks the startup social contract and messes with how families price the risk of joining early. Think the Windsor style deal. Forty in the room. Two hundred out of it.
George: The tension is real. I am sympathetic to rewarding the LeBrons, but companies are built by teams. We need mechanisms that share upside. Equity, bonuses, internal mobility. There is a good podcast on the sports culturification of engineering. Superstar comps, team dynamics, the whole thing. Reward the stars and the system that makes stars possible.
Tom: Give builders a short list for the next quarter.
George: One, design for flex. Separate latency sensitive inference from shiftable jobs. Make everything snapshotable. Two, make energy a feature. Do grid aware scheduling and build utility partnerships into your infra roadmap. Three, run dual regimes. Deterministic rails for transactions and probabilistic lanes for cognition, wrapped in continuous QA. Four, go geo modular. Localize models, data paths, and audit trails. Assume fragmented compliance. Five, mind the culture. If you chase superstars, codify broad upside so the rest of the team stays bought in.
Tom: Here is the through line. Winners will not just be model clever. They will be power literate, policy literate, and product disciplined. And they will still ship.
George: Could not agree more.
If this sharpened your map, subscribe for more straight talk on AI, markets, and the systems that actually make them work.
Keep building. Keep cranking.
My Stack:
Let’s Remember Jane Goodall as the Disruptor She Was
Jane Goodall died this week. She will be remembered as a gentle, matronly icon of conservation. I suspect the level of disruption she represented in the scientific community will be conveniently overlooked. A woman in an almost exclusively male-dominated field, to be sure. But more than that, she was an outsider who threw out the rule book entirely. Arriving in Gombe without a PhD, she broke every convention: naming chimps instead of numbering them, describing their emotions, documenting tool use, hunting, and even warfare. The establishment bristled — critics accused her of being unscientific, even unserious — but the data held, and the field bent to her. Precisely because she wasn’t trained in the “proper” way, she saw what the insiders overlooked. Cambridge gave her a doctorate after she’d already blown the doors off primatology. Let that be a lesson to anybody over-investing in pedigree.
Disinformation Is Breaking Up America’s Couples
Most of us know someone — or know someone who knows someone — who fell down the QAnon rabbit hole during COVID and never climbed back out. Families lost parents, kids, entire relationships. A new study from the University of Illinois puts numbers to the anecdotes, showing how misinformation and disinformation do more than fracture our democracies, they fracture our marriages. As the researchers put it: “They failed, at least in part, because those differing beliefs were associated with different realities that disrupted a shared identity and shared reality with their partner.” Mind your media diet accordingly.
Sources: University of Illinois
YouTube Folds to Trump to the Tune of $22 Million
“What’s the point of having fuck you money if you never say fuck you?” Priceless!