Loading...
Loading...
Loading...
Loading...
Loading...
Loading...
1 selected
Loading...
0 selected
Loading...
0 selected
0 selected
All post types✓
Portfolio updates✓
Stock lists✓
Research✓
News✓
All Content✓
Source feeds
Buzzberg Top 50
✓
All market capsNo capitalization filter✓
200 B and aboveMega✓
10 B to 200 BLarge✓
2 B to 10 BMid✓
0 to 2 BSmall✓
Custom
Enter market cap range in B USD
All directions✓
▲ Long✓
▼ Short✓
⛔ Avoid✓
✂ Close✓
◦ Others✓
Any score✓
LOW+✓
MED+✓
HIGH✓
15:57
Aug 25
Aug 25
Dwarkesh Patel
18d
SMH
META 1ST
ASML
MU
FXI
▾
HIGH
AI infrastructure capex boom benefits semiconductors
AI infrastructure capex is exploding from over $1 trillion this year to more than $2 trillion by 2028 and roughly $11 trillion from 2024-2029; every layer of the semiconductor and data center supply chain is being forced to rebalance and raise prices as compute demand outstrips supply, making the AI semiconductor supply chain a core beneficiary.
SMH LONG
Meta's compute optionality is undervalued
Meta is one of only two plausible third compute owners because it builds compute speculatively on its balance sheet without needing an end customer, giving it optionality to monetize internally or lease/sell to Anthropic/OpenAI at very high margins; at roughly $1.5 trillion, Meta is worth far more given its cash flows and hoarded compute infrastructure.
META LONG
EUV tool scarcity creates pricing power
EUV lithography is a severe bottleneck: a single EUV tool can be bought for roughly $400 million and resold for over $1 billion because AI compute demand is so extreme, while Carl Zeiss is only targeting about 100 tools/year and supply-chain expansion takes years, giving ASML/EUV tools massive scarcity pricing power.
ASML LONG
Memory pricing power is surging now
AI demand has shifted value capture toward memory: SK Hynix, Micron and Samsung can raise prices quickly, HBM/memory supply is tight and the business is going to do great, even if an AI-driven market regime should keep multiples at only 2-3x earnings rather than driving another 10x re-rating.
MU LONG
005930.KS LONG
000660.KS LONG
285A.T LONG
China semiconductor capacity inflects upward 2028
China has been held to under 10% of incremental AI compute by export controls, but domestic fabs from SMIC and CXMT start reaching millions of units per year by 2028, and China's manufacturing scale and subsidies mean its AI compute capacity will hockey stick, potentially adding 50 GW in 2029, even if the chips are lower quality.
FXI WATCH
0981.HK WATCH
Rising rates break fragile sovereign borrowers
AI compute capex is forcing hyperscalers and supply-chain players to issue trillions in debt, crowding out other borrowers and raising interest rates globally; countries with high debt, low tax revenue and frequent rollovers such as Pakistan and Nigeria are likely to default, similar to the 1980s Volcker shock.
Pakistan AVOID
NGE AVOID
Banks suffer from credit spread blowups
If AI capex debt issuance pushes credit spreads wider, banks are structurally vulnerable because their liabilities reprice faster than their assets, causing them to lose large amounts of money; they are a casualty of the same higher-rate regime.
KBE AVOID
Rising discount rates crush long-duration equities
Higher discount rates from AI-driven debt issuance will crush the present value of long-duration cash-flow equities even if the S&P 500 overall holds up; Johnson & Johnson, railways and Berkshire-type stocks with 30-year cash flows become unattractive because investors will demand much higher returns.
JNJ AVOID
BRK.B AVOID
HIGH
17:33
Aug 07
Aug 07
Dwarkesh Patel
1mo
Leading AI labs
AMZN 1ST
GOOGL 1ST
Large organizations
▾
HIGH
Deployment flywheel widens leading AI labs' edge.
Dwarkesh argues that when deployment becomes part of training, the leading AI model benefits from more users doing complicated work and giving feedback that can be integrated beyond the session window, making it smarter; therefore the returns to being ahead in the AI race accelerate in favor of the leading AI labs.
Leading AI labs LONG
Cloud switching costs sustain high margins.
The cloud provider analogy shows lock-in economics: cloud margins are high because switching clouds is time-consuming and expensive, and Amazon or Google's quarterly earnings reflect that they are doing just fine.
AMZN LONG
GOOGL LONG
Inference batching favors large organizations.
Once personalized weight forks require full weight updates, efficient inference requires batching thousands of concurrent sequences; a large company with many employees and agents can serve its weight fork efficiently, while an individual user at batch size one suffers more than two orders of magnitude worse efficiency, so personalized AI economics favor big organizations.
Large organizations LONG
HIGH
17:36
Aug 03
Aug 03
Dwarkesh Patel
1mo
AI compute
H100 GPU compute
GB200/GB300 GPU capacity
TSM 1ST
ASML 1ST
▾
HIGH
Compute prices can rise tenfold.
Anthropic's revenue is growing 10x per year while lab compute grows only 3x per year, and labs want to keep most compute for training rather than become inference cloud providers. Since >90% margins on intelligence seem unlikely to persist without being competed away, the main escape valve is higher compute prices. Spot compute is already up more than 40% since February, frontier secured GPU capacity rents near 2x spot, and a human-level software engineer on one H100-equivalent could generate over $250K per year, which is more than 15x the current H100 spot price.
AI compute LONG
H100 GPU compute LONG
GB200/GB300 GPU capacity LONG
TSMC AI wafer capacity hits saturation.
AI is absorbing leading-edge wafer allocation that previously went to smartphones and PCs; at TSMC's leading-edge N3 nodes AI will go from 60% to 86%. Once all leading-edge wafer capacity is absorbed by AI, that reallocation source of compute growth hits a wall, making TSMC's leading-edge capacity increasingly scarce.
TSM LONG
ASML EUV bottleneck constrains fab buildout.
Of the 3x yearly compute growth, 1.2x comes from building new fabs, and that process is bottlenecked through 2030 and potentially beyond by the ability to build ASML EUV machines. This gives ASML a structural bottleneck as AI compute capacity scales.
ASML LONG
HIGH
16:30
Jun 30
Jun 30
17:17
Jun 19
Jun 19
Dwarkesh Patel
2mo
Expert data labeling and RL environment industry
Humanoid robotics
▾
HIGH
AI training data industry will surge.
AI model progress is mainly driven by adding more and better data and by scaling compute, with RL functioning as synthetic data generation. The expert data labeling and RL environment industry is already earning billions in annual revenue and is set to reach deca-billions as models require bespoke human expert trajectories across every skill.
Expert data labeling and RL environment industry LONG
Human sample efficiency would unlock robotics.
If AI models could learn physical tasks as sample-efficiently as humans, robotics would become a deca-trillion-dollar industry and produce an army of humanoid robots such as Unitree G1s doing useful work. Current models remain too sample-inefficient and need millions of hours of demonstrations, so this is a conditional but important setup to monitor.
Humanoid robotics WATCH
HIGH
17:54
Jun 16
Jun 16
18:14
Jun 09
Jun 09
Dwarkesh Patel
3mo
ERUS 1ST
FXI 1ST
Container Shipping
▾
HIGH
Russia's continental model destroys wealth.
Russia lacks the prerequisites for a successful maritime prosperity model: no moat, many hostile neighbors, poor internal transport, unreliable warm-water sea egress, no commerce-driven economy, and unstable government institutions. Its continental security model relies on territorial conquest, creates failing buffer states, and destroys wealth in negative-sum wars, making Russia structurally unattractive as a broad investment exposure.
ERUS AVOID
China's continental turn is economically costly.
China shares the continental handicap: no defensive moat, many neighbors, narrow island-cluttered seas that become wartime kill zones, limited arable land and food import dependence, and Xi Jinping's shift toward crony-sector privilege over the private sector. Even though China benefited from the maritime trading order under Deng Xiaoping, its current continentalist turn and economically inferior Belt and Road land corridor make the policy path financially costly.
FXI AVOID
Sea freight beats land transport economics.
Sea freight and container shipping are structurally advantaged over land-based transport. Containerization reduced loading costs from nearly $6 per ton to under 20 cents, standardized containers fit trucks, rail, and ships, and the largest container ships can carry more than 21,000 containers with cargoes valued over $1 billion. Sea transport is far cheaper and more secure than rail corridors such as Belt and Road, which face unstable territories, multiple rail gauges, and chokepoints.
Container Shipping LONG
HIGH
16:37
Jun 04
Jun 04
Dwarkesh Patel
3mo
Relational sector
AI compute
DTCR 1ST
SPY 1ST
US Housing
▾
HIGH
Human-in-loop services stay scarce and valuable.
Even after broad automation, goods and services where a human in the loop is part of the product's value will remain scarce. Experiments show people pay more for human-made art because they value empathy, connection, and intrinsic human involvement, so value should accrue to the relational/human-intrinsic sector.
Relational sector LONG
AI compute demand may keep rising.
The historical Moore's-law pattern where computation value halved as supply exploded may be breaking: H100 rental prices are higher than three years ago because smarter models raise the opportunity cost of compute. If new uses keep appearing, compute demand may not satiate and its share of the economy could keep increasing.
AI compute LONG
Data center returns are currently super high.
Returns to data centers are currently very high. Whether that persists depends on whether capital owners satiate; if capital satiates, returns to accumulation fall and consumption rises, but for now data center returns are high.
DTCR LONG
Developing countries should index AGI gains.
For countries outside the AI production chain, Phil prioritizes indexing into AGI/global equity returns via sovereign wealth funds or equivalent over relying on retraining, because AI could arrive quickly. In an electricity-like scenario, broad index exposure captures the gains, and even with concentration it is worth trying to index now.
SPY LONG
US housing is poor AI wealth exposure.
Most households' capital is tied up in a house, but housing is uniquely ill-suited to be complementary to AI or robot production. Its current value is mostly land near other humans and relational factors, which will not be the main factor of production in an AI economy.
US Housing AVOID
Software demand is highly elastic.
Software is described as a particular kind of good with highly elastic demand: as it gets cheaper, people want disproportionately more of it, unlike oil, insulin, or agriculture where satiation limits total spending. That means software spending can keep expanding rather than automatically collapsing.
IGV LONG
HIGH
16:11
May 22
May 22
Dwarkesh Patel
3mo
NVDA
▾
HIGH
Nvidia B300+ FP4 speedup supports low-precision trend
Reiner argues that lower-precision AI arithmetic benefits from quadratic scaling in bit width, so halving precision should give more than a 2x throughput gain. He notes Nvidia historically doubled FLOPs when halving precision up to B100/B200, but in B300 and beyond Nvidia's product specs now show FP4 running at 3x the FP8 rate, reflecting that stronger low-precision advantage.
NVDA WATCH
HIGH
16:20
May 15
May 15
17:09
May 08
May 08
17:20
Apr 29
Apr 29
Dwarkesh Patel
4mo
HBM 1ST
NVDA 1ST
GOOG 1ST
▾
HIGH
Memory/HBM remains critical AI bottleneck.
Reiner reinforces the memory wall thesis: hyperscaler capex on memory is enormous, memory is a huge constraint for AI buildouts, and HBM bandwidth is the critical bottleneck for frontier inference, long context, and latency. He adds HBM is not getting hugely better, implying the bottleneck and pricing power persist.
HBM LONG
Nvidia rack-scale NVLink unlocks AI scale-up.
Nvidia's rack-scale NVLink/NVL72 architecture gives every GPU all-to-all connectivity within one rack, matching Mixture-of-Experts expert parallelism. Reiner argues one rack bounds the size of an expert layer, so larger scale-up domains are a huge unlock for bigger sparse models, lower weight-loading latency, and longer context; he credits Nvidia with a genuine ~4x scale-up increase via difficult rack design.
NVDA LONG
Google's large scale-up domains helped Gemini.
Google deployed very large scale-up domains for a long time while others were limited to smaller domains. Reiner says this hardware/scale-up lead explains why Gemini seemed to have successful pre-training for longer than some other labs, giving Google's AI infrastructure an early advantage for very large or sparse models.
GOOG LONG
HIGH
16:42
Apr 15
Apr 15
Dwarkesh Patel
5mo
CoreWeave 1ST
TSM 1ST
MU 1ST
SMH 1ST
NEBIUS 1ST
▾
HIGH
Nvidia-backed neoclouds benefit from GPU support.
Nvidia intentionally supports AI neoclouds such as CoreWeave, Nscale, and Nebius because converting AI CapEx into rentable compute is needed and Nvidia does not want to become a cloud itself; Jensen says these companies would not exist or be where they are without Nvidia support, and they are now doing fantastically.
CoreWeave LONG
NEBIUS LONG
TSMC critical AI foundry and packaging partner.
TSMC is scaling CoWoS and future packaging at the same level as logic, is the critical foundry for Nvidia's leading-edge nodes, and is unique among foundries in its ability to support Nvidia-scale demand and multi-decade trust; Jensen says no other foundry in history can offer that dependability.
TSM LONG
Micron benefits from HBM memory demand.
Micron was an early partner that believed Jensen's AI capacity forecasts and doubled down on LPDDR and HBM memory; Jensen says this has been tremendous for Micron as HBM has moved from specialty to mainstream computing technology.
MU LONG
China chips advance despite export controls.
Jensen argues China already has massive compute, abundant energy, roughly half of AI researchers, and substantial chip manufacturing capacity; Huawei just had its largest year and China's domestic AI/chip ecosystem will keep advancing despite export controls, so US policy cannot stop Chinese AI hardware progress and may accelerate it.
SMH LONG
Energy is AI's long lead-time bottleneck.
Jensen says chip and advanced packaging capacity bottlenecks clear in two to three years, but downstream energy is the harder constraint because AI factories, reindustrialization, EV manufacturing, and robotics all require large energy supplies, and energy projects take much longer; this makes energy a key bottleneck and sustained demand area.
XLE LONG
AI agents multiply software tool demand.
Jensen expects the number of AI agents and tool users to grow exponentially, which will cause instances of software tools such as Synopsys Design Compiler and other design, floor-planning, layout, and design-rule-check tools to skyrocket; software tool makers like Cadence and Synopsys benefit because agents are not yet good enough to use their tools, and the likely outcome is a combination of tool companies building agents and agents learning to use the tools.
SNPS LONG
CDNS LONG
Nvidia moat durable; ecosystem and TCO lead.
Jensen argues Nvidia sits at the center of the electrons-to-tokens transformation, which is hard to commoditize, and that its moat comes from the CUDA ecosystem, hundreds of millions of installed GPUs, presence in every cloud, best performance per dollar and per watt, scaled supply chain commitments, and an annual architecture roadmap such as Hopper to Blackwell improving efficiency 30x to 50x.
NVDA LONG
HIGH
16:33
Apr 07
Apr 07
16:26
Mar 13
Mar 13
Dwarkesh Patel
6mo
ASML 1ST
005930.KS 1ST
000660.KS 1ST
NVDA 1ST
CHIK 1ST
▾
HIGH
EUV bottleneck gives ASML pricing power.
By 2028-2029 the ultimate bottleneck for scaling AI compute shifts to ASML EUV lithography tools. ASML can produce roughly 70 EUV tools this year, 80 next year, and only a bit above 100 by end of decade; each gigawatt of Rubin requires about 3.5 EUV tools. No competitor has anything close to EUV and ASML has not raised prices as fast as tool capability, leaving large pricing power and upside in a capacity-constrained market.
ASML LONG
Memory prices surge as supply lags.
Memory is a major bottleneck: DRAM/HBM/NAND prices are inflecting up, memory vendors are raising prices again, and roughly a third of Big Tech capex may go to memory. No meaningful new fabs arrive until late 2027 or 2028 because memory makers underinvested after losing money in 2023, while AI long context and KV cache demand is absorbing capacity and forcing consumer demand destruction.
005930.KS LONG
000660.KS LONG
MU LONG
Nvidia locked up logic and memory.
Nvidia has locked up a majority of leading-edge logic and memory supply: it is TSMC's largest customer, is getting roughly 70 percent of N3 wafer capacity by 2027, has signed long-term contracts, and has been far more aggressive than Google or Amazon in signaling demand across PCB, memory, and packaging supply. This lets Nvidia capture margin in a compute-constrained world, and the utility of its installed H100s is rising rather than depreciating.
NVDA LONG
China scales indigenous chip production.
China is aggressively building an indigenous semiconductor supply chain and Dylan is quite bullish over five to ten years. By 2030 China should have fully indigenized DUV tools and working EUV tools, with more engineers and state capital than the West, and could scale chip production even if process technology lags. On long AI timelines, China's vertically integrated scale advantage grows.
CHIK LONG
Power equipment orders surge, supply scales.
Power is not the ultimate scaling bottleneck. The gas turbine and power equipment complex has huge data center orders; GE Vernova, Mitsubishi, and Siemens turbine capacity has been locked up by early movers who can now charge excess returns, and behind-the-meter gas, aeroderivatives, ship engines, and reciprocating engines can together add hundreds of gigawatts by the end of the decade.
ENR.DE LONG
7011.T LONG
GEV LONG
Memory costs squeeze Apple's iPhone margins.
Apple is squeezed by the memory and NAND price spike because iPhone bill-of-materials costs are rising by roughly $100 to $150 and consumers may not fully absorb higher prices. At the same time, Apple is becoming a smaller part of TSMC and is being crowded out of leading-edge capacity by AI/HPC chips, reducing its favorable supply-chain position.
AAPL AVOID
Memory costs crush low-end phone makers.
The memory cost spike is crushing low- and mid-range smartphone makers with thin margins and no long-term memory agreements. Xiaomi and Oppo are already cutting low-end and mid-range smartphone volumes by half, and smartphone volumes could fall from about 1.1 billion to 500 to 600 million next year as memory costs consume the bill of materials.
1810.HK AVOID
Google woke up to AI demand.
Google initially under-allocated AI compute and sold TPU capacity to Anthropic before Gemini demand inflected. Gemini reached about $5 billion ARR in Q4 and management is now aggressively buying energy companies, turbine deposits, powered land, and utility agreements. This AI awakening supports a much more aggressive Google AI buildout.
GOOG LONG
TSMC sold out, can prebook capacity.
TSMC is sold out on leading-edge logic for 2026, can only offer incremental 5 to 10 percent to latecomers like Google, and is prioritizing the higher-margin HPC/AI market over mobile. As AI demand explodes, TSMC can force customers to prebook and prepay for capacity while Apple becomes a smaller and less important customer.
TSM LONG
Bloom fuel cells scale fast.
Bloom Energy fuel cells are a behind-the-meter power solution for data centers. SemiAnalysis has been positive on Bloom for a year and a half because its production payback period is very fast and it can increase production quickly, even though fuel cells are more expensive than combined-cycle gas turbines.
BE LONG
HIGH
18:57
Mar 11
Mar 11
Dwarkesh Patel
6mo
AMZN
NVDA
GOOG
PLTR
▾
MED
Big Tech may drop Pentagon for AI.
The Department of War's supply-chain restriction against Anthropic creates a developing risk/setup for Big Tech defense contractors. If upheld, Amazon, Nvidia, Google, and Palantir would need to ensure Anthropic is not touching their Pentagon work. As AI becomes embedded across all products, it may become impossible to cordon off Claude use from defense work, forcing these companies to choose between their AI provider and a Pentagon relationship that is only a tiny fraction of revenue.
AMZN WATCH
NVDA WATCH
GOOG WATCH
PLTR WATCH
MED
Choose Sources
Loading sources...