Modest Proposal - AI Commoditization and Capital Dynamics - [Invest Like the Best, EP.380]
What this covers
The host talks with Modest Proposal, an anonymous investor who oversees a large pool of public and private capital, in their third conversation. The organizing idea is the capital cycle, illustrated with concrete cases. Beverage can makers Ball and Crown raised annual capex from roughly $600 million and $500 million to about $1.5 billion each to serve pandemic demand, then watched demand revert. U.S. freight shows the same pattern, with overcapacity still weighing on pricing two years after retailers stopped flowing goods. He also concedes that some distortion is useful, since mid-2000s commodity prices got mines built.
The AI portion applies this lens. He likens today's capex to roughly $120 billion spent on fiber in the late 1990s, and he doubts NVIDIA can stay the choke point for the largest companies, while treating ASML as a more durable case of process power. He lays out a coalescing view that LLMs commoditize as Llama and Mistral fast-follow, which would push surplus toward applications. On power, he weighs a forecast of about 2.5 percent annual growth after twenty flat years against efficiency gains, citing DeepMind at Google and flat data center consumption from 2015 to 2019.
The last stretch turns to allocation. The index has beaten most active investors over five and ten years, which he reads as exceptional earnings power in U.S. mega-caps rather than a bubble. He argues the endowment shift of 30% to 45% into privates has gone too far, and that private equity's mid-teens IRRs overstate investor returns given a roughly $300 billion gap between cash returned and invested. He also covers Japan's Tokyo Stock Exchange pressure on capital management plans, and behavioral adherence limits on GLP-1 drugs.
Modest argues that COVID-era demand shocks and AI-driven demand surges create temporary surpluses that markets wrongly capitalize as permanent, so investors should track capital cycles and barriers to new supply rather than chase the consensus winners or the index.
- Surging demand meeting inelastic supply produces temporary surpluses that build capacity, which later collapses into overcapacity and falling prices (commodities 2000s, beverage cans and freight post-COVID).
- Public indexes, especially the U.S. mega-caps, have been very hard to beat, and the same pressure has pushed capital into illiquid private markets that may be overallocated.
- AI value capture depends on whether frontier models stay differentiated or commoditize, which determines whether surplus accrues to infrastructure or applications.
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
As money flows out of active management, the remaining managers are not demonstrably outperforming, so the paradox of skill implies that the remaining skilled active managers face a harder environment rather than easier inefficiencies.
“I think it actually probably means it's harder.”
Beverage can makers Ball and Crown raised capex from about $600 million and $500 million a year to about $1.5 billion each when customers demanded capacity for pandemic-level demand of 10% a year, but demand reverted and left overcapacity.
“capital expenditures go from $600 million a year to $1.5 billion a year for each of the company.”
The key question in surplus situations is whether barriers to entry let a company keep the rapidly rising surplus, and markets price this uncertainty when they put rising multiples on rising surpluses.
“whether it's Helmer, Porter or Mauboussin how much of the rapid increase in surplus is a company able to hold on to.”
Once supply built during a demand-driven boom comes online, it is permanent, so commodity indexes remain depressed even after the demand that justified the investment fades.
“You look at commodity indexes still down, and that's not surprising because the supply is permanent.”
When surging demand meets inelastic supply, markets put high multiples on the temporary surplus, capitalism responds by building new supply over three to five years, and the surplus eventually disappears as supply comes online.
“when surging demand meets inelastic supply? And the mid-2000s commodities just did a wonderful example of that”
ASML is a choke point for the entire semiconductor process, and its position reflects process power, though smart people keep trying to dislodge it, and the question is duration rather than possibility.
“ASML is an incredible company. They're a choke point for the entire semiconductor process.”
The index has been undefeated over the past five and ten years, and very few active public investors are pleased with their relative performance, with the difficulty intensifying in the last five years.
“the index over the last five and 10 years has been undefeated.”
Spending roughly $120 billion on fiber optics in the late 1990s without a clear profit justification is comparable to today's AI capex, so the AI spend not yet being justified by revenue is not unusual for an infrastructure platform shift.
“we're putting $120 billion in fiber optics into the ground. What is the profit justifying this? Everybody would have laughed.”
Markets must sometimes capitalize temporary surplus, because if commodity prices had not risen in the mid-2000s, many mines would not have been built, so some distortion is needed to fund technological innovation.
“We probably do need the distortive impact in order to draw in the technological innovation that furthers society.”
GLP-1 benefits are extraordinary, but the main barrier to societal benefit is behavioral: adherence to chronic drugs falls off sharply, with cancer drug adherence around 60% to 70% and rapid drop-off after that, so sustained dosing is hard.
“the adherence to cancer drugs was something like 60% or 70% and then it rapidly fell off from there.”
The endowment model's shift of 30% to 45% of pool assets into illiquid privates has gone too far, and capital is moving there mainly for return smoothing and volatility avoidance rather than risk-adjusted return gains.
“the endowment model and the allocation of assets into privates, I think, has gone too far”
The Tokyo Stock Exchange is shaming companies into publishing forward capital management plans, and those that do not are listed publicly, which pushes Japanese firms with overcapitalization, cross-shareholdings, and low payouts toward shareholder-friendly plans.
“the Tokyo Stock Exchange essentially shaming companies into adopting more forward-looking capital management program.”
The index has been undefeated over five and ten years, and COVID amplified strong management and advantaged companies while stressing marginal companies and managements, so most active investors trail the benchmark.
“the index over the last five and 10 years has been undefeated.”
Pandemic-era demand for hard goods led to a surge of trucks and capacity in U.S. freight, and after retailers like Target and Walmart stopped flowing goods in 2022, the resulting overcapacity still persists two years later, keeping pricing weak.
“In May of 2022 Target and Walmart came out and said, "We are absolutely stuff to the gills with inventory."”
Private equity's claimed mid-teens IRRs overstate investor returns because there is roughly a $300 billion deficit of money returned versus money invested over three years, which hurts continuously compounded returns for investors.
“I saw a chart the other day that showed there's about a $300 billion deficit in money back versus money in over the last three years.”
A coalescing investment view holds that LLMs will become relatively undifferentiated commodities, with open source models like Llama or Mistral fast-following, so frontier models will be relegated to leading-edge niches and will not clearly outperform.
“LLM models are, if not commodities are going to be relatively undifferentiated and that the frontier models will push forward.”
If open-source models like Llama or Mistral become good enough and frontier models are undifferentiated, enormous surplus accrues to the application layer while the infrastructure captures only part of it.
“in a world with relatively undifferentiated models and open source being good enough, enormous surplus accrues to application”
The idea that one company can be the choke point for five of the largest companies in history is doubtful, because startups are likely working on ways to run compute more efficiently than GPUs, such as stripping functionality to run on CPUs.
“Do you think four guys in a garage in California aren't currently thinking of ways to more efficiently run the compute that is currently only efficiently done via GPUs?”
Assuming AI power demand will persist unchanged is a bet against human ingenuity, since DeepMind helped Google cut data center power use and data center consumption was flat from 2015 to 2019, so efficiency gains are likely.
“We know that DeepMind in the 2010s was able to help Google reduce power consumption in traditional data centers.”
U.S. power demand, flat for 20 years, is likely to grow sharply in the next two to three years because of AI, and if it does, grid investment will raise prices and expose end users to higher costs, which will change their behavior.
“2.5% a year growth over the next seven, eight years, and that summed up to over 30% aggregate growth in power consumption in the U.S.”
After 20 years of flat US power consumption, a report expects about 2.5 percent annual growth over the next seven to eight years, summing to over 30 percent aggregate growth, and Modest's gut says the growth rate will be right over the next two to three years because these shifts happen quickly.
“they expect something on the order of 2.5% a year growth over the next seven, eight years, and that summed up to over 30% aggregate growth in power consumption in the U.S.”
The sustained US tech and mega-cap outperformance is less a bubble than a reflection of exceptional compounding earnings power, so for the foreseeable future these companies remain some of the most incredible businesses and an active manager faces a tough hurdle.
“for the foreseeable future, you just have to say these are some of the most incredible businesses ever and they're still performing.”
Japan is a potentially interesting place for capital because the Tokyo Stock Exchange is pressuring companies to publish capital management plans, which could fix chronic issues like cross-shareholdings, low payout ratios and rare buybacks, although international investors have been burned before.
“the Tokyo Stock Exchange essentially shaming companies into adopting more forward-looking capital management program.”
Markets repeatedly put multiples on temporary surplus created when surging demand meets inelastic supply, implicitly capitalizing a long duration of cash flows that the surplus will not sustain, and this is a constant source of confusion for the speaker.
“the way I think of it in microeconomic terms was temporary surplus. And when you put a multiple on something, you're implicitly capitalizing a long duration of cash flows.”
Investors can find attractive setups by catching troughs in capital cycles, buying when companies have overspent and then pulled back, because the capex trough often precedes a new up cycle.
“if you can catch the trough of those, they overspend then they pull back, then capital spending trough and then you get a new up cycle.”
Mean reversion arguments that mega-cap tech must revert have been wrong for years, and the companies are in no worse shape than five years ago, with some in stronger positions, so the speaker views the end as unpredictable in timing.
“people have been calling for this end for years and years. And I think the companies are in no worse shape today than they were five years ago”
If frontier models stay significantly better than open source for most tasks, surplus accrues to Google, Microsoft, OpenAI, and AWS and hurts many software companies, whereas commoditized models push surplus up to the application layer.
“incredible surplus accrues to Google, to Microsoft and OpenAI, AWS, big loser, lots of software companies, big loser.”
Buying NVIDIA for a 12-month return requires believing frontier AI labs will spend far more on compute, while justifying the current price over the long run requires believing a single company can remain the choke point for the economy's largest companies against new efficient approaches.
“you have to differentiate between are you putting 10% of your portfolio in NVIDIA because you think it's going to be higher in 12 months or because you think it's going to sustainably generate the surplus”
The more interesting AI change is architectural shifts in products and markets, which are happening in software but at companies too small to matter yet in public markets.
“what's far more interesting is if there are architectural shifts in products and markets... those are going to be tiny companies today”
Displacing Apple would be extremely hard because the smartphone is the unifying device of every ecosystem, and succeeding would likely require a model powerful enough to abstract away integrated hardware and software, which means believing in AGI.
“displacing Apple, you almost have to believe in AGI. You have to believe in a model so powerful that it basically abstracts away the need for the integrated hardware software solution.”
Private market returns are largely a reflection of public markets, so a substantial decline in forward public 10-year returns would also drag private market returns, and most late entrants will get Nasdaq-like returns with illiquidity.
“to some extent, actually, to a large extent, the private markets are a reflection of the public market.”
Meta spending heavily on fast-following open-source models could commoditize its complement, and data gravity means models will run where enterprise data already sits, such as AWS or Azure.
“if Meta's going to spend $40 billion fast following, they're going to commoditize their complement and then you'll just run these models where your data is and data gravity is real”
Investors who joined private markets late are accessing non-persistent managers and taking on illiquidity for probably Nasdaq-at-best returns, while early investors' few persistent venture funds are immaterial to their overall asset base.
“you are getting access to the nonpersistent performers and you are now exposing yourself to probably Nasdaq at best returns with an illiquid profile”
Mega-cap multiples in the low-20s to low-30s are elevated versus a 75-year average of mid-to-high teens, but they mainly reflect the compounding of earnings power rather than irrational pricing.
“It's more a reflection of the fact that their earnings power over that entire period of time, has just grown.”
Bonds now offer real yield for the first time in a long time, with the environment potentially providing 1% to 2% real returns on the 10-year Treasury, which is a very different environment from the previous 15 years of near-zero rates.
“today, you actually have bonds or an option for the first time in a long time, so you can earn yield.”
The speaker's best buys were typically based on a known controversy or tension whose resolution was expected in the near term, with the investor taking the other side of consensus, while sells have diverse causes like plan completion, error, or better opportunities.
“the ones that worked out the best, I think you would be able to recreate, "Okay, XYZ is the story."”
The investment community has written off the scenario in which a single generalizable model becomes so good it usurps specialized models, even though that outcome would be very different from the commoditized world, and no one knows which path is correct.
“it's interesting to me that the investment community has written that path off.”
Market structure now features a much higher threshold for being overlooked, with stocks below roughly $15 billion market cap often ignored, and momentum persisting longer because of algorithmic trading, making markets more trending.
“the definition of what I would call a smaller, less liquid name has moved from a $1 billion market cap to $15 billion”
The size at which a stock becomes neglected has risen from about a $1 billion market cap to about $15 billion, leaving a wide range of $5 to $10 billion companies without news-driven coverage that get left behind.
“the definition of what I would call a smaller, less liquid name has moved from a $1 billion market cap to $15 billion”
Adding up current AI software applications yields roughly mid-single-digit billions of revenue, which is far short of the roughly $200 billion in capex, so the revenue side alone does not justify the spending.
“you can add up all the cats and dogs software applications, and you'd probably get to mid-single-digit billion.”
AI productivity and efficiency gains on the cost side are real but invisible in revenue numbers, so a full economic accounting must aggregate cost avoidance alongside revenue to judge whether AI spending pays off.
“What you can't really see is the cost avoidance.”
If AI is only a sustaining innovation, incumbents will add a copilot and charge customers about $20 more, making the shift technologically boring rather than transformative.
“every incumbent slaps a copilot into their product, then they charge you $20 more, and this is actually a really boring technological shift.”
If GLP-1 drugs take a large share of the country out of obesity and diabetes, the reduced long-term costs to Medicare, Medicaid, and insurance would free productive resources at a very large scale.
“the economic benefit of taking a huge chunk of the country out of obesity, out of diabetes, out of the long-term costs”
Markets repeatedly keep putting multiples on temporary demand surges, including the 2020 surge in digital demand, which was unlikely to persist unless the real world shut down permanently.
“Soaring demand for everything digital in 2020 when the real world shut down, that probably wasn't going to persist”
A software investor should ask what the alternative is when Microsoft has compounded earnings at mid-teens for ten years, since that is a high bar for any other holding.
“if my opportunity cost is Microsoft growing mid-teens, what am I doing here?”
Early venture investors with a few persistent top funds have an immaterial share of their asset base, so even extraordinary MOICs have diminishing impact on overall returns for large pools.
“those 5 or 10 funds probably represent an immaterial portion of their overall asset base.”
The stress of COVID and its aftermath magnified strong management and competitive advantages while exposing marginal companies and managements, bifurcating the market.
“the stress of COVID and everything that's transpired since really magnified strong management and competitive advantages and really stressed marginal companies”
The shortage-then-glut pattern is only broken where a barrier prevents new supply from coming online, so the investable question is identifying which markets have those barriers.
“that's the back that markets make when they put rising multiples on rapidly rising surpluses.”
Higher real interest rates and a depressed housing market, with about 4 million existing homes sold a year versus roughly 5.5 million normal, create a rubber band that could snap back in residential, homebuilding, private equity, and commercial real estate once front-end rates ease.
“we're selling four million existing homes a year versus should be probably 5.5.”
Frontier AI labs that believe they are near AGI will spend more than most people expect, which means NVIDIA will sell more GPUs than consensus expects, and the speaker thinks a paper on AGI timelines hints at these spending intentions.
“Those labs seemingly are going to be spending bigger sums than I think most people realize or expect because they fervently believe they're on the cusp of building AGI.”
The best reading of AI investment is that frontier labs will grow compute intensity by 100x to 10,000x, and that trajectory, not stock price momentum, is what must be believed for AI capex to pay off.
“the frontier labs want to spend in a way to grow their compute intensity 100x, 1,000x 10,000x, and that seems to be the trajectory that they are on.”
The Nasdaq 100 has delivered about 5x MOIC on a rolling 10-year basis, yet very few venture fund families have produced multiple funds at 5x, so public indexes are beating venture at persistence.
“how many fund families do you know of that have multiple 5x MOICs?? It's not many.”
China joining the WTO and going from a very small to a very large consumer of globally traded commodities within five or six years pushed commodity prices up sharply, which drew capital into miners and emerging market indexes until supply eventually came online and prices collapsed by 2014.
“China had historically been out of the global economy for a long time, joined WTO... went from consuming a very small amount of global traded commodities to a very large amount.”
Aggregates (rocks) are hard to disrupt because quarries are local monopolies due to transport distance, so a cheaper synthetic material would not be enough to displace them.
“They're local monopolies effectively, the distance to transport the rocks is the challenge and they own the rocks and you don't.”
Amazon expanded its distribution network so much in two years that it effectively rebuilt UPS's capacity, and then suffered the hangover when demand slowed.
“Amazon built UPS in two years. That's the quoted stat”
The equal-weighted S&P 500 beat the market-cap-weighted index on a rolling 10-year basis from about 2000 to 2018, and the cap-weighted index then recaptured its lead, now running about 2.8% a year over 10 years.
“the equal-weighted was actually ahead on a rolling 10-year basis consistently from basically 2000 to 2018”
Oil fell from about 140 to about 80 and commodity indexes stayed down after 2014 because the supply added during the boom was permanent, even as China's growth slowed.
“oil was, it was 140 and now it's 80”
The move from on-premise software to SaaS was a disruptive architectural shift, and the open question now is whether AI-native companies with real-time queryable data architectures will do the same to SaaS.
“on-prem to SaaS was a pretty disruptive architectural shift, and I think the question today is, are there AI native companies”
A key consumer question is whether horizontal search remains the downstream gatekeeper or gets verticalized into specialized agents, though it is too early to know.
“does a horizontal search type product remain the gatekeeper downstream? Does that get verticalized with specialized agents”
Google's search profit pool once looked impossible to attack, but since Perplexity is raising money, the venture market no longer treats search as untouchable, a shift from the signal of enormous power.
“I think clearly, that's no longer the case since Perplexity is raising money.”
A large pool considering venture should size the allocation so it can matter to total return, and check how much capital can actually be deployed into top-quartile venture funds.
“think about the size of that venture portfolio to matter to your overall return stream, then think about how much that capital is as a percent of top quartile venture funds”
For a large allocator investing internationally, identifying the opportunity and then finding the right manager to capture it is very hard, and picking managers is probably as hard or harder than picking individual securities that outperform.
“picking managers is probably as hard or harder than picking individual securities to outperform.”
Surging demand for AI chips meets inelastic supply, since TSMC sells as fast as it can make chips, making semiconductors a prime example of the surplus dynamic.
“chips where as fast as TSMC can make chips, they are being sold.”
Many great non-U.S. companies, including European ones, are usually priced at or above their comparable U.S. peers, so they are not cheap, though some international exposure is still warranted in case U.S. outgrowth fails.
“they are usually priced at or above their comparable U.S. peer.”
Prices in motion tend to keep moving in the same direction more vigorously and longer than before, plausibly because of algorithmic trading, making trends more pronounced until they stop.
“objects in motion stay in motion longer, meaning whether it's algorithmic trading or what have you.”
Stories of an AI chip company reaching $200 billion in sales at 80% gross margins will draw attention, but they are future speculation rather than a current reason to buy the stock.
“it's going to be $200 billion in sales at a 80% gross margin, some crazy number, generally, that will draw a lot of attention. But that's in the future.”
Ira Sohn conference pitches have shifted from distressed debt and turnarounds toward technology-focused ideas like ASML and AI companies, reflecting where returns have been.
“this year, we had Daniel Gross interviewing the CEO of magic.dev. We had ASML winning the Best Ideas contest.”
Two people who recently left or work within OpenAI publicly predicted AGI within three to five years, while others predict it will not arrive, and the divergence matters for how the AI investment landscape shakes out.
“two people, one who's left OpenAI and one, two nights ago within OpenAI, basically say, we think AGI is here within three to five years.”
The Klarna story about AI productivity is probably apocryphal, though some real productivity and efficiency gains are occurring on the cost side.
“The Klarna story is probably apocryphal.”
SaaS went from the best business model in the world to companies growing 10% to 12% with no margins in roughly three years, showing how fast the environment shifts.
“SaaS is the best business model in the world three years ago and now everyone is wondering”
Freight has not yet reached a capital cycle inflection, and people have been trying to call the bottom for a year without success.
“Candidly, freight is not there yet. People have been trying to call this for a year.”
Warren Buffett's three I's framework is a relevant frame, though the speaker did not elaborate on what it implies in this context.
“Warren Buffett famously talks about the three I's.”
Attacking a profit pool like aggregates is not worth a talented entrepreneur's effort because better uses exist than displacing rock, reinforcing that barrier-heavy niches are low-priority targets.
“if you're smart enough to do that, there's probably some other better use than displacing rock.”