Wall Street is still underpricing the AI build-out: the notes put the spend at $7.6T over 5 years, but NVDA and AVGO just printed earnings that imply much larger demand. The map is a five-layer stack—energy, chips, infrastructure, models, applications—where power is the bottleneck, NVDA and AVGO are sold out, TSM, MU, SK Hynix, ANET, VRT, LTE, COHR, FN, POWL and the utilities sit in the middle, and GOOGL and META own the top layer that monetizes it. The call is to buy the winners at each layer and dollar-cost average because the biggest AI labs are still private, cash-burning, and mostly not directly investable yet.
Wall Street is pricing $7.6T of AI spending over the next 5 years, but NVDA and AVGO earnings already imply the number is too low.
NVDA printed record quarterly revenue of $96.2B, up 106% y/y; data center revenue was $89B, up 117% y/y.
AVGO printed record quarterly revenue of $29.6B, up 86% y/y; AI revenue grew 221% y/y.
AVGO said 73% of AI revenue came from custom chips/ASICs for six AI giants, including OpenAI, Anthropic, GOOGL, and META; 27% came from networking.
NVDA's AI mix is about 80% compute and 20% networking, including Spectrum-X Ethernet, Quantum InfiniBand, and NVLink.
NVDA guided to 94% AI revenue growth y/y next quarter; AVGO guided to 236% y/y next quarter.
Over the past year NVDA added $48B of quarterly AI revenue, more than 4x AVGO's $11.5B increase.
Goldman Sachs' chart assumes global AI spend grows 32% next year, 21% in 2028, then 4% by 2031, while NVDA says the top five cloud companies will spend over $800B this year and $1.3T next year.
Google and Amazon have raised CapEx budgets mid-year, sometimes more than once, because each new tranche of capacity immediately sells out.
The buildout is a five-layer cake: energy, chips, infrastructure, models, applications; AI is measured in tokens, not cached like the internet, because intelligence is produced in real time.
Power is the bottleneck: US utilities are planning for roughly 90GW of new data-center demand by 2030, about 18GW of always-on load per year, while the grid adds a record 86GW this year mostly from solar, wind, and batteries that only count for about 25% of sticker rating as always-on power.
The grid has been retiring more always-on coal and gas power than it has been adding, more power is waiting for grid connection than the whole grid has today, median queue time is over 5 years, and electricity prices have to rise to clear the gap.
CEG runs America's largest nuclear fleet and is restarting the healthy reactor at Three Mile Island on a 20-year Microsoft deal; VST locked 20-year nuclear deals with Amazon and META; GEV's gas turbines are effectively sold out for years.
NVDA's AI factories imply about $18B per GW for Hopper, $25B per GW for Blackwell, and $40B per GW for Vera Rubin; 1 week later Hock Tan said AVGO's chips and networking average $20B-$30B per GW.
Vera Rubin is already in full production, expected to be the fastest product ramp in NVDA's history, already set to be 20% of NVDA data center revenue next quarter, and a major driver of the 70% full-year growth NVDA guided.
AVGO has six major custom chip customers; it co-designed Google's TPUs and META's inference chips, and META's inference chips enter production this quarter.
Anthropic is scaling from 1GW of custom AVGO chips this year toward 10GW by 2028; OpenAI's first dedicated chip, jalapeno, just started shipping.
Hock Tan says AVGO already secured enough chip capacity to reach $115B of AI revenue next year and $230B the year after, implying a 4x AI revenue increase in 2 years if he's right.
NVDA GPUs and AVGO ASICs are not direct competitors: Google, META, Anthropic, and OpenAI are buying both at gigawatt scale, using GPUs for speed/flexibility and ASICs to cut costs on predictable, high-volume inference workloads like Search, Facebook/Instagram, ChatGPT, and Claude.
TSM makes chips for both NVDA and AVGO and now gets about 2/3 of revenue from AI and high-performance computing; MU and SK Hynix make the HBM all these chips need, with both companies' revenues more than tripling y/y.
NVDA's Ethernet revenue grew about 160% y/y; AVGO's AI networking business grew about the same, and its Tomahawk Ultra switches push Ethernet inside the rack against NVLink after AVGO taped out the industry's first 200 Tbps switch.
ANET expects over $3.6B of AI revenue this year; VRT handles power and cooling inside AI data centers and just raised guidance across the board.
In March NVDA invested $2B each into LTE and COHR, both laser/transceiver names; NVDA's purchase commitments, plus the fact that the AI build-out is adding optical ports faster than co-packaged optics are consolidating lasers, argues lasers sell for years.
FN assembles optical modules for NVDA, Amazon, and Cisco, but only has 9 analysts; POWL builds industrial switchgear, has a market cap under $7B, backlog up 69% y/y, and only 4 analysts.
Google alone now processes 3.2 quadrillion tokens per month, equal to reading a billion books a day, about 8x all the books humanity has ever written.
OpenAI and Anthropic filed paperwork to go public this summer; Anthropic could list as early as next month, and until then GOOGL and META are the only directly investable frontier-model builders.
Microsoft still owns a large chunk of OpenAI, and Amazon's Anthropic stake generated over $50B of paper gains last quarter alone.
META's ad engine may be the biggest AI application on Earth: AI-driven targeting lifted ad revenue 27% y/y and advertisers paid 12% more per ad.
Nearly 1 in 2 people on Earth already use GOOGL's business, and its AI search features reach billions of users monthly, giving the top layer a huge base to monetize.
The portfolio map at the end is CEG, VST, GEV for power; NVDA, AVGO, TSM for chips; COHR, LITE, FN, and POWI for infrastructure; GOOGL and META at the top; the thesis is to own the winners at each layer and dollar-cost average.
GOOGL is up 45% over the last year, more than 2x the S&P 500, yet the notes still frame it and META as the names to pound on the table.
The sponsor is Emergent, pitched as a full-stack AI builder that turns a plain-English prompt into a working product on a real URL in 11 minutes, with frontend/backend together and self-fixing behavior.
The speaker says Emergent was used to manage a spreadsheet of about 40 AI and chip stocks, with sorting by distance from the 52-week high and then by revenue growth, and says 'I own it.'