TickerTain — Make Analysis Fun Again

TickerTain

←

▶ 3:46 · AI Infrastructure · Frontier labs to control majority of world's incremental compute by end of 2025

episode shorts · 7

How Elon Played the Compute Market - Dylan Patel

The Export Controls Are Working - Dylan Patel

We're Worried About the Wrong Kind of Centralization - Dylan Pa…

Two Labs Are About to Take Half the World's New Compute - Dylan…

Why Isn’t China Further Behind in AI? - Dylan Patel

AI Doesn't Need to Be Good at Politics to Change Everything - R…

Why AI labs are shelving their best models - Dylan Patel

now playing · AI Infrastructure

AI Infrastructuretailwindscore 9/10dylan patel

Frontier labs to control majority of world's incremental compute by end of 2025

OpenAI and Anthropic will consume 40-50% of new compute in 2025 and over half by late 2025, driven by revenue per megawatt of $50-100M that lets them outbid all other customers for scarce c…

AI Infrastructuretailwindscore 9/10dylan patel

Frontier labs to control 70-80% of world's incremental compute by 2028

Anthropic and OpenAI are accelerating their share of global incremental compute from 30% in 2024 to 40-50% in 2025 and 70-80% by 2028, driven by superior revenue per megawatt ($50-100M+) th…

AI Hardware & Chip Architecturemixedscore 7/10dylan patel

Custom silicon (Google TPU, OpenAI chips) reducing lab dependence on Nvidia merchant GPUs

Frontier labs are vertically integrating into chip design (OpenAI) and securing dedicated TPU capacity (Anthropic via Google) to bypass Nvidia pricing and ensure supply priority for trainin…

AI Infrastructuretailwindscore 9/10dylan patel

Frontier labs to control 70-80% of world's incremental compute by 2028, driving prices to $50M/megawatt

OpenAI and Anthropic's revenue per megawatt ($50-100M+) dwarfs all other compute consumers ($10-15M), creating a bidding war that centralizes >50% of new compute to two labs by end of 2025…

Semiconductorsheadwindscore 8/10dylan patel

EUV mirror bottleneck limits leading-edge wafer supply through 2030 regardless of capital

Carl Zeiss mirror production for ASML EUV tools caps leading-edge fab expansion at ~100 tools/year by 2030, creating a hard physical constraint that cannot be solved quickly even with unlim…

Critical Minerals & Supply Chainriskscore 8/10dylan patel

EUV mirror bottleneck (Carl Zeiss) is the hardest constraint; $400M tool flips to $1B+ in secondary market

Mirror production for ASML EUV tools has the longest lead time and least elasticity in the semiconductor supply chain; even unlimited capital cannot accelerate output before 2027-28.

Grid & Power Infrastructuretailwindscore 7/10dylan patel

Turbine and transformer shortages creating arbitrage opportunities in power equipment

Long-lead-time power equipment (turbines, transformers) has become a bottleneck asset class, with buyers reselling at large premiums to data center developers desperate for energized capaci…

Data Center Infrastructuretailwindscore 9/10dylan patel

Compute pricing inflecting from $10-15M to $25-50M/megawatt as labs outbid commodity users

Frontier labs' ability to generate $50-100M+ revenue per megawatt will force compute pricing up 3-5x from current $10-15M/megawatt levels, as labs must outbid commodity inference providers…

Memory & Storagetailwindscore 8/10dylan patel

Memory suppliers raising prices fastest, capturing disproportionate AI value vs foundries

HBM and DRAM suppliers (Micron, SK Hynix, Samsung) are increasing prices far more aggressively than TSMC, shifting value capture up the stack as memory becomes the critical bottleneck for A…

Semiconductorstailwindscore 8/10dylan patel

Memory suppliers (Hynix, Micron, Samsung) capturing more AI value than foundries or chip designers

HBM/DRAM is the tightest bottleneck in the AI hardware stack; memory companies are raising prices 'very quickly' while TSMC raises 'very slowly', shifting value capture up the supply chain…

Neoclouds & Cloud Computingtailwindscore 7/10dylan patel

SpaceX, Fluidstack and specialized neoclouds becoming compute landlords auctioning capacity to highest-bidding labs

Entities with balance sheets (Meta, SpaceX) or specialized ops (Fluidstack) build speculatively then rent to OpenAI/Anthropic at $25-50M/megawatt, creating a new asset class between hypersc…

Neoclouds & Cloud Computingtailwindscore 8/10dylan patel

Merchant compute providers (SpaceX, CoreWeave-style) capturing scarcity rents by building speculatively

Entities with balance sheets (SpaceX, Meta) or specialized neoclouds can build compute without pre-signed contracts, then auction capacity to labs at 3-4x standard rates, creating a new pow…

AI Economics & Business Modelstailwindscore 8/10dylan patel

Labs will shift compute from inference to training as internal R&D returns exceed external inference revenue

When inference generates $60-70M/megawatt, labs still allocate more to training because the discounted cash flow of achieving AGI first dwarfs inference profits; inference fraction is alrea…

AI Economics & Business Modelstailwindscore 9/10dylan patel

Labs will redirect compute from inference to training as internal R&D returns exceed external token revenue

When revenue per megawatt reaches $60-100M, frontier labs maximize long-term value by allocating incremental compute to automated AI research rather than serving external customers, reversi…

AI Economics & Business Modelstailwindscore 9/10dylan patel

Frontier labs will shift compute from inference to training as internal R&D returns exceed external token sales

As revenue per megawatt rises from $50M to $100M+, labs maximize value by allocating marginal compute to internal AI research (automated coding/research) rather than selling inference, beca…

AI Geopolitics & Export Controlstailwindscore 8/10dylan patel

China's AI compute to hockey-stick in 2028-29 via SMIC/CXMT domestic fabs despite 2-3x quality deficit vs US chips

Export controls have held China to <10% of incremental compute, but massive state subsidies will drive 5-10 GW domestic additions in 2028 and potentially 50 GW in 2029, though quality-adjus…

AI Geopolitics & Export Controlsmixedscore 8/10dylan patel

China's AI compute to remain <10% of global incremental through 2027, then hockey-stick in 2028-29

Export controls have restricted China to sub-10% of new AI compute deployments, but domestic fab ramp (SMIC, CXMT) and potential policy shifts could enable 50 GW/year by 2029, though chip q…

Data Center Infrastructuretailwindscore 9/10dylan patel

AI CapEx trajectory implies $10T/year by 2030 — 10% of global GDP — requiring massive credit expansion

Current $2T/year CapEx growing to $10T by 2030 as labs triple compute yearly; power plants (30-yr assets) and datacenters (20-yr) must be built years ahead, pulling forward trillions in cre…

Data Center Infrastructureheadwindscore 8/10dylan patel

Power and construction lead times make 100 GW/year build rates physically implausible

Building 100 GW of annual AI compute requires power plants and data center shells started years in advance, creating a hard infrastructure ceiling that capital cannot immediately overcome.

Grid & Power Infrastructuretailwindscore 9/10dylan patel

Power plant and data center lead times make $10T annual CapEx by 2030 a physical reality

Building 100+ GW/year requires power plants (30-year assets) and data centers (15-20 year assets) to be constructed years ahead, pushing total annual CapEx toward $10T by 2030 — ~10% of glo…

Macro & Ratesheadwindscore 9/10dylan patel

AI-driven credit demand will trigger second Volcker shock, defaulting emerging markets and crushing non-AI equity valuations

$5T+ of new AI infrastructure debt will raise spreads 250+ bps, forcing 40%+ of tax revenue to interest in debtor nations and repricing all long-duration equities (utilities, consumer stapl…

Macro & Ratesriskscore 9/10dylan patel

AI-driven credit demand will push corporate borrowing rates to 8%+, triggering sovereign debt crises in emerging markets

Hyperscalers and labs willing to pay 8%+ for debt to fund compute will raise spreads economy-wide, causing a Volcker-style shock where 40+ developing nations default and non-AI equities re-…

Macro & Ratesheadwindscore 8/10dylan patel

AI-driven rate surge risks sovereign debt crises in emerging markets and crushes value-stock multiples

A 250-500bps rise in long-term rates from AI capital demand would push 60%+ of US tax revenue to debt service, trigger emerging-market defaults (Volcker shock repeat), and collapse DCF valu…

AI Bubble / Capex Debateriskscore 9/10dylan patel

$11T AI CapEx 2024-2029 requires $5T new debt, driving structural interest rate increase

The $11T cumulative AI infrastructure buildout (2024-2029) requires $5T in new credit issuance even after maximum cash-flow funding, creating a structural demand for capital that will push…

AI Infrastructuretailwindscore 9/10dylan patel

AI CapEx to exceed $2T annually by 2028 with $11T cumulative through 2029

Global AI infrastructure spending is on track for $2T+ per year by 2028, requiring $5T+ in new credit issuance that will structurally raise interest rates and crowd out non-AI investment.

AI Bubble / Capex Debateriskscore 8/10dylan patel

$11T cumulative AI CapEx (2024-29) requires $5T debt funding; revenue per megawatt must keep compounding to avoid bust

SemiAnalysis models $11T total buildout with only $6T cash-funded; the $5T credit gap assumes labs' revenue/megawatt keeps rising — if regulation or model stagnation stalls monetization, th…

Public Markets & Valuationmixedscore 7/10dylan patel

In high-growth AI regime, all equities should trade at 2-3x earnings as discount rates converge to GDP growth rate

If economy doubles yearly (100% growth), risk-free rate approaches tens of percent; DCF models collapse for non-AI stocks, making even AI hardware (Micron, Kioxia) trade at trough multiples…

Public Markets & Valuationriskscore 7/10dylan patel

If AI transforms economy, all equities should trade at 2-3x earnings — memory stocks included

In a true AI takeoff scenario where interest rates reach tens of percent, the discount rate forces all long-duration cash flows (including memory semiconductors) to trade at 2-3x earnings;…

←

▶ 3:46 · AI Infrastructure · Frontier labs to control majority of world's incremental compute by end of 2025

episode shorts · 7

How Elon Played the Compute Market - Dylan Patel

The Export Controls Are Working - Dylan Patel

We're Worried About the Wrong Kind of Centralization - Dylan Pa…

Two Labs Are About to Take Half the World's New Compute - Dylan…

Why Isn’t China Further Behind in AI? - Dylan Patel

AI Doesn't Need to Be Good at Politics to Change Everything - R…

Why AI labs are shelving their best models - Dylan Patel

now playing · AI Infrastructure

AI Infrastructuretailwindscore 9/10dylan patel

Frontier labs to control majority of world's incremental compute by end of 2025

OpenAI and Anthropic will consume 40-50% of new compute in 2025 and over half by late 2025, driven by revenue per megawatt of $50-100M that lets them outbid all other customers for scarce c…

AI Infrastructuretailwindscore 9/10dylan patel

Frontier labs to control 70-80% of world's incremental compute by 2028

Anthropic and OpenAI are accelerating their share of global incremental compute from 30% in 2024 to 40-50% in 2025 and 70-80% by 2028, driven by superior revenue per megawatt ($50-100M+) th…

AI Hardware & Chip Architecturemixedscore 7/10dylan patel

Custom silicon (Google TPU, OpenAI chips) reducing lab dependence on Nvidia merchant GPUs

Frontier labs are vertically integrating into chip design (OpenAI) and securing dedicated TPU capacity (Anthropic via Google) to bypass Nvidia pricing and ensure supply priority for trainin…

AI Infrastructuretailwindscore 9/10dylan patel

Frontier labs to control 70-80% of world's incremental compute by 2028, driving prices to $50M/megawatt

OpenAI and Anthropic's revenue per megawatt ($50-100M+) dwarfs all other compute consumers ($10-15M), creating a bidding war that centralizes >50% of new compute to two labs by end of 2025…

Semiconductorsheadwindscore 8/10dylan patel

EUV mirror bottleneck limits leading-edge wafer supply through 2030 regardless of capital

Carl Zeiss mirror production for ASML EUV tools caps leading-edge fab expansion at ~100 tools/year by 2030, creating a hard physical constraint that cannot be solved quickly even with unlim…

Critical Minerals & Supply Chainriskscore 8/10dylan patel

EUV mirror bottleneck (Carl Zeiss) is the hardest constraint; $400M tool flips to $1B+ in secondary market

Mirror production for ASML EUV tools has the longest lead time and least elasticity in the semiconductor supply chain; even unlimited capital cannot accelerate output before 2027-28.

Grid & Power Infrastructuretailwindscore 7/10dylan patel

Turbine and transformer shortages creating arbitrage opportunities in power equipment

Long-lead-time power equipment (turbines, transformers) has become a bottleneck asset class, with buyers reselling at large premiums to data center developers desperate for energized capaci…

Data Center Infrastructuretailwindscore 9/10dylan patel

Compute pricing inflecting from $10-15M to $25-50M/megawatt as labs outbid commodity users

Frontier labs' ability to generate $50-100M+ revenue per megawatt will force compute pricing up 3-5x from current $10-15M/megawatt levels, as labs must outbid commodity inference providers…

Memory & Storagetailwindscore 8/10dylan patel

Memory suppliers raising prices fastest, capturing disproportionate AI value vs foundries

HBM and DRAM suppliers (Micron, SK Hynix, Samsung) are increasing prices far more aggressively than TSMC, shifting value capture up the stack as memory becomes the critical bottleneck for A…

Semiconductorstailwindscore 8/10dylan patel

Memory suppliers (Hynix, Micron, Samsung) capturing more AI value than foundries or chip designers

HBM/DRAM is the tightest bottleneck in the AI hardware stack; memory companies are raising prices 'very quickly' while TSMC raises 'very slowly', shifting value capture up the supply chain…

Neoclouds & Cloud Computingtailwindscore 7/10dylan patel

SpaceX, Fluidstack and specialized neoclouds becoming compute landlords auctioning capacity to highest-bidding labs

Entities with balance sheets (Meta, SpaceX) or specialized ops (Fluidstack) build speculatively then rent to OpenAI/Anthropic at $25-50M/megawatt, creating a new asset class between hypersc…

Neoclouds & Cloud Computingtailwindscore 8/10dylan patel

Merchant compute providers (SpaceX, CoreWeave-style) capturing scarcity rents by building speculatively

Entities with balance sheets (SpaceX, Meta) or specialized neoclouds can build compute without pre-signed contracts, then auction capacity to labs at 3-4x standard rates, creating a new pow…

AI Economics & Business Modelstailwindscore 8/10dylan patel

Labs will shift compute from inference to training as internal R&D returns exceed external inference revenue

When inference generates $60-70M/megawatt, labs still allocate more to training because the discounted cash flow of achieving AGI first dwarfs inference profits; inference fraction is alrea…

AI Economics & Business Modelstailwindscore 9/10dylan patel

Labs will redirect compute from inference to training as internal R&D returns exceed external token revenue

When revenue per megawatt reaches $60-100M, frontier labs maximize long-term value by allocating incremental compute to automated AI research rather than serving external customers, reversi…

AI Economics & Business Modelstailwindscore 9/10dylan patel

Frontier labs will shift compute from inference to training as internal R&D returns exceed external token sales

As revenue per megawatt rises from $50M to $100M+, labs maximize value by allocating marginal compute to internal AI research (automated coding/research) rather than selling inference, beca…

AI Geopolitics & Export Controlstailwindscore 8/10dylan patel

China's AI compute to hockey-stick in 2028-29 via SMIC/CXMT domestic fabs despite 2-3x quality deficit vs US chips

Export controls have held China to <10% of incremental compute, but massive state subsidies will drive 5-10 GW domestic additions in 2028 and potentially 50 GW in 2029, though quality-adjus…

AI Geopolitics & Export Controlsmixedscore 8/10dylan patel

China's AI compute to remain <10% of global incremental through 2027, then hockey-stick in 2028-29

Export controls have restricted China to sub-10% of new AI compute deployments, but domestic fab ramp (SMIC, CXMT) and potential policy shifts could enable 50 GW/year by 2029, though chip q…

Data Center Infrastructuretailwindscore 9/10dylan patel

AI CapEx trajectory implies $10T/year by 2030 — 10% of global GDP — requiring massive credit expansion

Current $2T/year CapEx growing to $10T by 2030 as labs triple compute yearly; power plants (30-yr assets) and datacenters (20-yr) must be built years ahead, pulling forward trillions in cre…

Data Center Infrastructureheadwindscore 8/10dylan patel

Power and construction lead times make 100 GW/year build rates physically implausible

Building 100 GW of annual AI compute requires power plants and data center shells started years in advance, creating a hard infrastructure ceiling that capital cannot immediately overcome.

Grid & Power Infrastructuretailwindscore 9/10dylan patel

Power plant and data center lead times make $10T annual CapEx by 2030 a physical reality

Building 100+ GW/year requires power plants (30-year assets) and data centers (15-20 year assets) to be constructed years ahead, pushing total annual CapEx toward $10T by 2030 — ~10% of glo…

Macro & Ratesheadwindscore 9/10dylan patel

AI-driven credit demand will trigger second Volcker shock, defaulting emerging markets and crushing non-AI equity valuations

$5T+ of new AI infrastructure debt will raise spreads 250+ bps, forcing 40%+ of tax revenue to interest in debtor nations and repricing all long-duration equities (utilities, consumer stapl…

Macro & Ratesriskscore 9/10dylan patel

AI-driven credit demand will push corporate borrowing rates to 8%+, triggering sovereign debt crises in emerging markets

Hyperscalers and labs willing to pay 8%+ for debt to fund compute will raise spreads economy-wide, causing a Volcker-style shock where 40+ developing nations default and non-AI equities re-…

Macro & Ratesheadwindscore 8/10dylan patel

AI-driven rate surge risks sovereign debt crises in emerging markets and crushes value-stock multiples

A 250-500bps rise in long-term rates from AI capital demand would push 60%+ of US tax revenue to debt service, trigger emerging-market defaults (Volcker shock repeat), and collapse DCF valu…

AI Bubble / Capex Debateriskscore 9/10dylan patel

$11T AI CapEx 2024-2029 requires $5T new debt, driving structural interest rate increase

The $11T cumulative AI infrastructure buildout (2024-2029) requires $5T in new credit issuance even after maximum cash-flow funding, creating a structural demand for capital that will push…

AI Infrastructuretailwindscore 9/10dylan patel

AI CapEx to exceed $2T annually by 2028 with $11T cumulative through 2029

Global AI infrastructure spending is on track for $2T+ per year by 2028, requiring $5T+ in new credit issuance that will structurally raise interest rates and crowd out non-AI investment.

AI Bubble / Capex Debateriskscore 8/10dylan patel

$11T cumulative AI CapEx (2024-29) requires $5T debt funding; revenue per megawatt must keep compounding to avoid bust

SemiAnalysis models $11T total buildout with only $6T cash-funded; the $5T credit gap assumes labs' revenue/megawatt keeps rising — if regulation or model stagnation stalls monetization, th…

Public Markets & Valuationmixedscore 7/10dylan patel

In high-growth AI regime, all equities should trade at 2-3x earnings as discount rates converge to GDP growth rate

If economy doubles yearly (100% growth), risk-free rate approaches tens of percent; DCF models collapse for non-AI stocks, making even AI hardware (Micron, Kioxia) trade at trough multiples…

Public Markets & Valuationriskscore 7/10dylan patel

If AI transforms economy, all equities should trade at 2-3x earnings — memory stocks included

In a true AI takeoff scenario where interest rates reach tens of percent, the discount rate forces all long-duration cash flows (including memory semiconductors) to trade at 2-3x earnings;…