Rao warns AI's energy bill to exceed chip costs by 2027
- Energy now makes up 40% of AI compute costs, projected to surpass 50% by 2027, flipping the economics of scaling.
- OpenAI is booking power and chip capacity for 2030 and paying for its own grid infrastructure to secure gigawatt builds.
- The shift to power-hungry AI is breaking the Fed's ability to slow big tech with interest rate hikes.
The physical ceiling on artificial intelligence is no longer the chip. It’s the plug. Naveen Rao, a guest on This Week in AI, projects the energy portion of AI’s total compute cost will exceed 50% within four years. With the latest GPU clusters, power already accounts for nearly 40% of expenses, up from just 10% a generation ago.
OpenAI’s financial planning confirms the bottleneck has moved from fab to grid. On the All-In podcast, CFO Sarah Friar said the company is negotiating for compute capacity in 2030 and beyond. A one-gigawatt data center now carries a $50 billion price tag including land, power, and chips. To keep building, OpenAI is paying for its own power infrastructure for a Michigan project to avoid spiking local utility rates.
"We are hitting an energy wall that manufacturing cannot solve alone. Simply building more chips won't work if the grid can't light them."
- Naveen Rao, This Week in AI
Monetary policy is powerless against this capital cycle. On TFTC, analyst John Tinsman argued that hyperscalers like Microsoft and Google are spending trillions on compute infrastructure with returns on invested capital exceeding 35%. A one or two percent hike in interest rates is irrelevant when companies fund growth with cash and lease compute at 90% profit margins.
The scale of demand is breaking old models. Tinsman pointed to Elon Musk’s Colossus data center, built for $4 billion and leased to Anthropic for billions annually, as the new template. AI agents, unconstrained by human headcount, create insatiable token demand. Rao estimates 20-30% of current compute is wasted on “token maxing” to game corporate leaderboards.
Public sentiment is the final constraint. Rao and This Week in AI co-host Alex Finn argued that “doomer” narratives from firms like Anthropic have painted AI as a public threat, sparking local protests against data centers over water and power use. Their proposed fix: AI companies should fund tangible local benefits, like public transit, for communities hosting their infrastructure.
"If a data center uses a community's grid, the company should fund free public transit or better schools for that specific town."
- Alex Finn, This Week in AI
The race is no longer just for faster chips, but for intelligence per watt. Rao said the future belongs to non-Von Neumann architectures that mimic the brain’s efficiency. For now, the companies that secured power contracts years in advance hold the only viable scaling path.
Source Intelligence
- Deep dive into what was said in the episodes
AI Layoffs, Compute Costs & Agents | Naveen Rao & Alex Finn on This Week in AI Episode 16 • Jun 4
- Naveen Rao estimates that 20-30% of current AI compute token costs are wasted on 'token maxing,' a gaming of usage metrics driven by leaderboards and corporate proxy goals.
- Current AI models lack the holistic reasoning, architectural foresight, and production-grade reliability of a senior human developer. Alex Finn counters that the intelligence is already revolutionary; the problem is its misapplication by non-technical users.
- Alex Finn reports his coding velocity has increased by a thousandfold using AI. He attributes this to deeply understanding systems, not just prompt blasting.
- Alex Finn runs Quen 3.7 locally on a $4,000 Nvidia DGX Spark, advocating for 'unlimited, dumber intelligence' to power 24/7 agents for tasks like scraping social media for opportunities.
- Naveen Rao notes that the total cost of ownership for GPU clusters is shifting from capex to opex, with energy now constituting nearly 40% of TCO for current-gen Nvidia chips. He projects this will exceed 50% within the next 3-4 years.
- Naveen Rao blames 'doomer' narratives, specifically calling out Anthropic, for painting AI as an existential threat. He argues this damages public perception, fuels protests against data centers, and risks harmful regulation.
- Alex Finn traces current AI layoff rhetoric to irresponsible hiring during the 2020 zero-interest rate period. He argues CEOs are using AI as a scapegoat for prior overspending, not as the real cause of cuts.
- Naveen Rao identifies a core problem as Silicon Valley's failure to let the public share in AI's financial upside, exacerbated by companies staying private too long. He contrasts this with China, where public sentiment views AI as a competitive superpower.
- Alex Finn posits that seizing private equity for a public trust destroys incentives. He proposes a policy alternative: give every American a funded ChatGPT plan and education on extracting value from AI.
- Naveen Rao suggests AI companies building data centers should voluntarily invest in local communities, like funding public buses or rec centers, to build tangible public goodwill and counter misinformation-driven protests.
- Alex Finn and Naveen Rao both express skepticism about buying into imminent hyped IPOs like Anthropic or SpaceX, citing distorted valuations and a preference to let price discovery settle first.
Also discussed on this episode: (1)
AI Infrastructure (1)
- Naveen Rao's startup, Unconventional AI, is developing non-von Neumann architectures where memory and compute are unified. He aims for a 2-3 order of magnitude improvement in power efficiency to overcome the coming energy wall.
OpenAI CFO Sarah Friar: IPO, AI Rivalries, New Device, and Spending $100B+ on Compute • Jun 2
- OpenAI is developing a new consumer device with Jony Ive, described as natural, lovable, and intimate. Friar says it will be unveiled by year-end and available for purchase early next year.
- Friar sees ChatGPT as a hybrid of Google and Meta, possessing high user intent data plus memory and demographic context. This creates a potent ad platform, though an ad-free tier will remain.
- OpenAI’s current token allocation prioritizes strategy over pure economics; API tokens are an order of magnitude more valuable than consumer tokens, but they are provisioning for broad global access.
Also discussed on this episode: (10)
Big Tech (4)
- OpenAI completed a $120 billion fundraising round, which CFO Sarah Friar calls the most successful in history. The previous largest IPO was Saudi Aramco at roughly $30 billion.
- Friar frames an IPO as a fundraising milestone, not a destination, to create optionality. OpenAI raised $22 billion in March for flexibility.
- Friar describes OpenAI’s strategy as building the AI infrastructure layer with multiple interfaces. ChatGPT serves over 900 million weekly users and has become the noun and verb for AI.
- Usage intensity scales sharply with pricing tiers. Free users average about seven turns daily, the first paid tier doubles that to 15, the $20 Plus tier triples it, and Pro users see an 11x increase over free.
Enterprise (1)
- OpenAI’s revenue is now roughly balanced 50-50 between consumer and enterprise. Friar cites heavy enterprise engagement with firms like Thermo Fisher, major banks, Travelers, and tech companies.
AI Infrastructure (4)
- Compute scarcity is the current bottleneck, with insufficient tokens available through 2026 and limited supply in 2027. Friar credits OpenAI’s early compute buying for its current position.
- OpenAI’s capital strategy involves shifting capex to opex by partnering with multiple cloud providers. They now work with Oracle, CoreWeave, Microsoft Azure, Google Cloud, AWS, and smaller neoscalers.
- OpenAI is diversifying its chip strategy beyond Nvidia. Their fall training run will use Vera Rubens, AMD chips are in the pipeline, Cerebras is live for low-latency dev work, and they are developing their own chip with Broadcom.
- OpenAI’s new one-gigawatt data center in Celine, Michigan will create 2,500 union jobs, pay $1 billion in taxes, and invest $45 million in Codex education credits. Friar emphasizes community trust and not raising local electricity bills.
Models (1)
- Model efficiency gains are driving down customer costs despite rising compute input prices. GPT-4 to GPT-5 saw a 97% cost deprecation; GPT-5’s 2X price increase still yields a 20-30% lower cost per token for customers.
Marty Bent
#752: Why AI Stocks Are Cheap with John Tinsman • Jun 1
- Tinsman sees the US winning the AI compute race by building data centers domestically using cheap natural gas, then leasing compute globally at ~90% profit margins, bypassing expensive energy export costs.
- Software companies leveraging AI face unleashed demand; tools like Adobe see usage soar as the limiting factor shifts from expensive human operators to cheap AI agents.
- Tinsman is bullish on AI's net job impact, comparing it to the steam shovel at the Panama Canal: displaced workers find higher-value roles, increasing overall GDP and wealth.
- The fertilizer industry faces a supply crisis: strife in the Straits of Hormuz disrupts sulfur and helium, while natural gas shortages shut down global nitrogen plants, spiking input costs over 100%.
- Tough farm economics are emerging as fertilizer prices spike while crop prices lag; John Deere has tightened financing, pushing 40% of Tinsman's customers to non-traditional lenders at high rates.
Also discussed on this episode: (6)
Markets (1)
- John Tinsman's AOTG ETF selects stocks based on high earnings growth and low marginal cost, favoring businesses like Microsoft over capital-intensive ones like Boeing. This contrasts with market-cap weighted index funds.
AI Infrastructure (5)
- XAI's deal with Anthropic reveals extreme data center profitability: a $4B data center leased for $15B annually yields a 10x ROI in three years. Tinsman argues hyperscalers' true returns are a closely guarded industry secret.
- AI token demand is exploding, with Anthropic's tokens sold growing 80x in 12 months; Goldman Sachs projects another 24x growth. This dwarfs traditional market expectations for 20-30% revenue growth.
- Hyperscaler AI capex is becoming rate-insensitive; Morgan Stanley estimates it will reach $1.1T by 2027. With ROIs potentially near 100%, spending continues despite high interest rates.
- Elon Musk's Colossus 1 data center was built in 122 days using a bitcoin-mining style approach: daisy-chained generators and battery walls to shorten the distance between power and compute.
- The AI build-out is creating massive local economic impact; Tinsman notes Iowa welders and electricians building data centers now earn $250k annually, up from $50k.