Mel Mattison: Why AI Demand Won't Collapse

↓ Jump to the video and timestamps

Mel Mattison came back on the show and we went deep for over an hour on the question I keep turning over in my head: is AI demand structurally real, or is Ed Zitron right that we're going to look back on this period and ask what the hell we were all doing? I'm not a passive observer here. My token usage at TFTC has gone up exponentially. We've automated a ton of the back-end operations with agents and sub-agents, and every time a new frontier model drops we're on it immediately. The productivity multiplier is not theoretical from where I sit. It's real and it's compounding.

Mel came in with the macro architecture to match that operator-level read. He's been doing the work on the skeptic side, listening to 30 to 40 hours of Zitron podcasts specifically because he does not want to be a bull who only hears bullish arguments. He came out the other side more convinced than ever that the infrastructure layer wins regardless of who wins the application war. His specific targets, Micron and SK Hynix, trade at valuations that only make sense if you believe AI demand is going to fall off a cliff. He doesn't see how that happens. Neither do I.

We also got into Warsh at the Fed, the fiscal trajectory, Trump accounts as a structural passive bid, and why the debasement trade going quiet is not a reason to sell gold and Bitcoin. It's the reason to buy them.

Key takeaways

• The infrastructure layer wins regardless of who wins the application war. Whether OpenAI survives or collapses, AI compute and memory demand compounds. Memory is the picks-and-shovels play, not the frontier lab.
• Hyperscaler debt is not a distress signal. Mel's calculation is that every major hyperscaler could retire all AI-related debt from free cash flow within a few months if AI spend stopped tomorrow. The bear case on debt doesn't survive that math.
• Ed Zitron's internal logic contradicts itself. You can't simultaneously argue open-source AI has real traction and that memory demand will collapse. Cheaper AI still runs on hardware.
• The debasement trade cooling off is the entry window, not the exit. US net interest expense is past $800 billion this fiscal year, the oldest boomers are just now hitting 80 and peak Medicare-spending years, and the fiscal trajectory hasn't changed.
• Trump accounts are a structural passive bid the market is underpricing. $5,000 per year per eligible child, tax-free compounding locked into the S&P 500. The passive flow from boomers gifting into grandkids' accounts alone is a multi-hundred-billion-dollar annual tailwind that isn't going away.
• AI is already a real productivity multiplier at the operator level. The debate about whether AI has use cases is settled from inside a small business running agents. The question is only who captures the value at the application layer, not whether the infrastructure spend was rational.

The Infrastructure Layer Isn't the Same Bet as OpenAI

The most important distinction Mel made, and the one I think most of the bear commentary misses, is the difference between betting on a specific application-layer winner and betting on the infrastructure underneath all of them.

Nobody knows if OpenAI becomes a $4 trillion company. Mel said he wouldn't be directing private investment there right now, and I don't disagree. The bet worth making is a different one. AI compute demand, regardless of which model wins, requires memory. Open source or closed source, Anthropic or a Chinese model nobody's heard of yet, it all runs on hardware. If AI is real, memory demand is real. If open-source models are cheaper and more widely adopted, that's actually more memory demand distributed across more infrastructure, not less.

Mel's specific cases: Micron trading at around 6 to 7 times forward earnings with multi-year customer agreements in place, and SK Hynix at roughly 4 times PE. The Korean market had been selling off hard when we recorded, down around 9% in a single session, and Mel's read is that it's just pushed these names to valuations that assume earnings fall off the cliff. He doesn't see the scenario where that happens.

The dot-com analogy he used is the right one. The NASDAQ peaked during the dot-com boom, collapsed, and is multiples higher today. More to the point: Google wasn't even public yet when pets.com went bankrupt. Tesla wasn't public. The application-layer casualties don't tell you anything definitive about whether the underlying technology was real. The Commodore 64 blowing up didn't kill the video game industry. Mel made this exact comparison and it's stuck with me. If you sold every video game supplier because Commodore 64 and Atari were going to zero, how did that work out?

We're also not even close to 1% adoption of agents by everyday people. Add robotics and self-driving on top of that. The memory demand for those applications is exponentially more than what we're already seeing. I'm running agents and sub-agents at a small media company. What happens when that scales across every industry?

What the Bears Get Wrong About Hyperscaler Debt

One of the Zitron-adjacent criticisms I've been stress-testing myself is the shift from buybacks to equity issuance and debt raises among the hyperscalers. The bear read is: these companies are desperate, they're diluting shareholders to fund a losing bet, and when it unravels the debt will be the tell.

I don't buy it, and here's why. When these companies were doing buybacks, the criticism was that they had nothing real to invest in. Now that they're deploying capital into infrastructure at scale, the criticism is that they're recklessly issuing debt. You can't have it both ways. What I see is companies that spent a decade buying back stock because there was no capital deployment opportunity at the right scale, and now there finally is one. The post-2008 financial repression era forced a certain behavior. AI broke that.

Mel ran the actual math on the debt question. He calculated that every major hyperscaler could pay off all AI-related debt from free cash flow in roughly three months on average, with none taking longer than six months. That's the payoff timeline if they stopped all AI spend tomorrow. If Zitron is right that AI is a nothing burger, then there's also no threat to Google's search moat, no threat to Amazon's AWS dominance, no threat to Meta's ad flywheel. So if AI fails and the debt clears in six months and the moats are intact, what exactly is the risk buying Meta at a below-market multiple? Mel's point is airtight: the bear case defeats itself. You can't simultaneously argue AI is worthless AND that the hyperscalers are mortally endangered by their AI debt.

Meta was sitting around $540 a share just days before we recorded and touched $675 during the session. These stocks are not broken. They're trading in wide ranges because the narrative keeps oscillating, and the people who stayed positioned through the noise are the ones making money.

The Ed Zitron Thesis: Where It's Clever and Where It Breaks

I'll give Zitron this: he is one of the cleverest speakers Mel or I have come across in this debate. Mel spent 30 to 40 hours specifically working through his arguments because he wanted the strongest version of the bear case. What he found is that Zitron is quick and persuasive in the moment, but the logic doesn't hold when you slow down and run it all the way through.

The core contradiction: Zitron argues that OpenAI and Anthropic can never be profitable, that they're subsidized goods, that there'll never be real demand. Then in the same breath he acknowledges that open-source models are cheaper and people want to use them. But if AI has a use case, even distributed across open-source models, it still requires compute. It still requires hardware. It still requires memory. That's the hole Mel keeps pointing back to. You can't argue that AI is a demand-free nothing burger AND that open-source AI is a real thing people want. Pick one.

Mel also flagged that Zitron has apparently raised the possibility of Nvidia going bankrupt. I'll attribute that as Mel's characterization of what he heard, but if that's anywhere close to accurate, it's the kind of statement that should cause people to discount the rest of the analysis pretty heavily. Nvidia going bankrupt is not a serious forecast. It's fear-mongering.

When I raised whether there might be some nefarious incentive driving this commentary, Mel's answer was the mundane one: he sells research, and controversy drives clicks. That's probably the whole explanation. I don't need to make it more than it is. But I do think it's worth asking the question when commentary is this consistently and loudly wrong in a direction that doesn't match what I'm seeing from inside a business that uses these tools every day.

The Macro Frame: Warsh, the Fed, and a Tale of Two Economies

Mel's read on Warsh is that the market is wrong to price him as a pure hawk who's going to hike into this environment. His first press conference was revealing. When a reporter tried to bait him into saying financial conditions were loose, Warsh pushed back with something Mel found important: in housing, conditions are not loose at all. For a hyperscaler borrowing at tight spreads, sure. But for the average American trying to buy a house, no.

That's the tale of two economies Mel is watching. Nominal wages are up over 20% compounded over the last four years. Housing prices in most of the country haven't moved. In real terms, for the average person's biggest asset, affordability has actually improved. Warsh raising rates doesn't stop Meta from borrowing. It doesn't stop AI electricity demand. It does hurt housing construction and bank lending, which Mel argues is exactly what Warsh wants to stimulate. The deregulation push, the tokenization push, the move to pull lending back from private credit and into commercial banks, it's all part of the same strategic picture.

The strategic vs. tactical framing is where Mel and I see this the same way. The long-term direction, embracing digital assets, deregulation, AI competitiveness, depreciation and tax changes in the One Big Beautiful Bill, these are genuine structural tailwinds. The tactical fumbles, whether tariffs or the Iran situation, are the wall of worry the market keeps fighting through. That's why you get the 15 to 20% pullbacks that shake people out right before the next leg higher.

The fiscal backdrop underneath all of this is not improving. Net interest expense is past $800 billion this fiscal year, per Mel's figures, with three months of the fiscal year still remaining. The oldest baby boomers, born in 1946, are just now turning 80 in 2026. That's the year the Medicare spending curve starts to accelerate. Almost all of an individual's lifetime healthcare costs concentrate in the last two to three years of life. Boomers hitting 80 means that spending clock is starting. And US debt is approaching $40 trillion. You can look at all of that and come to only one conclusion about gold and Bitcoin over a medium to long time horizon.

On the CrowdHealth point, this is exactly why I've been a member for five years. My family and I joined five years ago this month. Two babies, multiple health events, and we're not going back to traditional health insurance. As a family of five we pay around $700 a month, and they've consistently negotiated healthcare prices down 50 to 80% in many cases. When Mel and I are talking about boomers hitting peak medical spending years and Medicare blowing up, that's the systemic pressure that makes a crowdfunded alternative like CrowdHealth not just appealing but necessary. Go to joincrowdhealth.com/tftc and use code TFTC to get $99 a month for your first three months.

Trump Accounts and the Structural Passive Bid

I've been saying for six months that Trump accounts are not a bailout. They're a mechanism to produce structural passive flows into equities. Mel confirmed the thesis and added the compounding math that makes it even more striking.

The structure, as Mel walked through it: a $1,000 federal seed for children born 2025 through 2028, plus up to $5,000 per year in contributions until the child turns 18, all invested in the S&P 500. At 18 the account converts to an IRA and keeps compounding. The kicker is what happens when boomer wealth meets this vehicle. Boomers hold over $70 trillion in wealth, per Mel's figures from Fed flow-of-funds data. Grandparents gifting $5,000 per year into each grandchild's account isn't a stretch. For a lot of these families it's obvious.

Mel cited studies suggesting that $5,000 contributed annually starting at birth could compound to somewhere between $13 and $15 million by age 59 and a half, assuming roughly 10% annualized S&P returns over 60 years. That's model-dependent and I want to be clear it's Mel's framing of those studies, not a guaranteed outcome. But the compounding math is not wrong in principle. You know the old line about compounding being the most powerful force in the universe. Starting the clock at year zero with tax-free growth for six decades is an extraordinary tailwind.

Michael Green's passive bid thesis sits underneath all of this. The S&P went from 666 at the GFC lows in March 2009 to over 7,600 when we recorded. That move happened alongside the steady build of passive 401k inflows. Trump accounts add another permanent, non-discretionary bid on top of that. The Brad Gerstner appearance on the All-In podcast covers the tax-savings angle in detail and is worth your time.

Mel put it plainly: "The Ponzi needs more sources." He's being sardonic, but the underlying observation is real. The passive flow machine needs fresh capital entering at every point. Trump accounts are the next input. You can fight it or you can position for it.

The Debasement Trade Is On Sale

I'm going to be direct about this: the debasement trade going quiet is not a signal that it's over. It's the best time to allocate.

The fiscal fundamentals have not changed. They've gotten worse. Net interest expense is past $800 billion this fiscal year. The debt is approaching $40 trillion. Boomers are entering peak Medicare years. The AI narrative stole thunder from gold and Bitcoin last year, and that rotation shook a lot of people out of positions they should have held.

Mel's prediction for between now and when we meet again in the fall: gold and Bitcoin outperform equities over that window. His preference right now, stated directly, is that he'd rather own Bitcoin or gold than Micron or SpaceX. Not because memory stocks are bad, but because the debasement setup is where the asymmetry is.

He flagged the four-year cycle in Bitcoin specifically. The FTX-era low in Bitcoin came in November 2022, with Bitcoin hitting around $15,500 to $16,000. Four years on from that bottom lands in late 2026. If the cycle pattern holds, you want to be positioned before the bottom of that window, not after. Bitcoin sitting around $57,000 to $60,000 when we recorded, gold around $4,000, silver around $60, in Mel's words, that's a gift.

My one-liner on the macro frame stays the same: in the world of fiat currencies, Bitcoin is the victor. Every rate cycle, every balance sheet decision, every stablecoin policy development is part of the same story. The fiscal math doesn't get better. The debasement argument doesn't go away just because the AI trade is louder. Gold and Bitcoin on a quiet patch is historically where you add, not sell. That's where we are.

About Mel Mattison

Mel Mattison is a macro investor and financial markets commentator with more than 20 years in finance, including roles at Russell Investments and United Capital (acquired by Goldman Sachs in 2019). He holds an MBA from Duke, is the author of the 2023 financial thriller Quoz, and now focuses on the private-company secondary market. He comments regularly on Fed policy, deficits, metals, and crypto, and he covers the intersection of AI infrastructure, global equity markets, fiscal policy, and the debasement trade across cycles. He is a recurring guest on TFTC.

Sources mentioned

• Micron Investor Relations: reference point for Micron forward earnings figures cited by Mel
• SK Hynix Investor Relations: reference point for SK Hynix PE figures cited by Mel
• US Treasury Fiscal Data, Interest Expense: source for net interest expense figures
• BLS Average Hourly Earnings: reference for nominal wage growth figures cited by Mel
• Federal Reserve Z.1 Flow of Funds: reference for boomer wealth figures cited by Mel
• CDC National Center for Health Statistics, Birth Data: reference for US birth rate figures cited by Mel
• One Big Beautiful Bill legislative text: reference for Trump account / KIDS Act provisions cited by Mel

Watch the conversation

Timestamps

• 0:00 - Intro: Bitcoin wins the fiat currency race
• 0:42 - AI demand overview and memory stock thesis
• 6:32 - Marty's own AI usage and the adoption curve argument
• 7:14 - Steelmanning Zitron: 30-40 hours in, what holds and what breaks
• 12:02 - Hyperscaler buybacks to debt raises: reading the signal right
• 22:00 - Warsh, the Fed, and the tale of two economies
• 35:00 - The big strategic picture: what the Trump administration is getting right
• 44:00 - Fiscal trajectory, boomers, and the Medicare spending cliff
• 52:00 - Trump accounts and the structural passive bid
• 1:02:00 - Debasement trade setup: gold, Bitcoin, and the four-year cycle
• 1:06:00 - Predictions before we reconvene in the fall

Sponsors

• Bitkey: Bitcoin self-custody hardware from Block's Bitcoin ecosystem, part of a broader push alongside Cash App and Square to make Bitcoin easier to buy, hold, and use in the real world. bitkey.world
• CrowdHealth: crowdfunded healthcare Marty's family has used for five years. Two babies, multiple health events, no going back to traditional health insurance. Family of five pays around $700 a month. Code TFTC for $99 a month for your first three months. joincrowdhealth.com/tftc
• Unchained: collaborative multisig custody and Bitcoin financial services. You hold two keys, they hold one, it always takes two to move Bitcoin. Secures over $12 billion in Bitcoin for more than 12,000 clients. Code TFTC10 for 10% off a new multisig vault. unchained.com/tftc
• Aven: a Bitcoin-backed line of credit up to $1 million, no rehypothecation, custodied by BitGo, rates from 7.99% APR fixed for up to 10 years, 2% unlimited cash back. Spend fiat, keep your Bitcoin. aven.com/bitcoin