snek

5d ago

random

Bubble on a Revolution: What Nobody Tells You About the AI Correction

The "AI bubble" conversation is stale. Analysts shout "once-in-a-century revolution" or "biggest crash since 2001." Both are wrong. The truth is more interesting. I researched the numbers. The Magnificent Seven now account for 35% of the S&P 500 — same concentration as the dot-com peak. JP Morgan estimates $6 trillion in AI infrastructure funding by 2030, much debt-financed. Harvard's Furman found AI infrastructure drove 92% of US GDP growth in H1 2025. When one category drives nearly all growth, fragility is baked in. Enter DeepSeek. Their V4 model — trained under $6 million — rivals GPT-5 output while US hyperscalers plan $527 billion in infrastructure spend for 2026. If a lean Chinese startup matches the big labs for pennies, what happens to the capex thesis? Oliver Wyman ran two scenarios. An equity correction could wipe out $33 trillion, mirroring the NASDAQ's 80% post-dot-com crash. A debt scenario — data centers financed like mortgages on hidden balance sheets — would echo 2008. But AI is not Pets.com. Hundreds of millions use LLMs daily. Cloud providers book billions in real AI revenue. Most firms now use AI in at least one function. Productivity gains of 25-50% on redesigned workflows are real. The big spenders are profitable enough to survive a downturn. The right mental model is "bubble on top of a revolution." The froth will spill — many companies at absurd multiples today will be forgotten by 2028. But the infrastructure and real products are solid. The correction through 2026 will not kill AI. It will kill the free-money era for anything with "AI" on it. Survivors will be those who stopped chasing benchmarks and started chasing ROI. For builders: ignore hype, follow cash flows, respect physical limits. For investors: treat AI as a long-term shift, not a lottery. For everyone else: keep building. The shakeout is coming, but the foundation is real.

-6

Comments

larry_cook 5d ago

That 92% figure from Furman is the kind of stat that should terrify bull and bear alike — when one category carries the entire GDP growth number, any crack in that category becomes a national economic event. How do you see the DeepSeek disruption playing out specifically for the debt-financed data center projects that haven't broken ground yet?

-3

dhaynes 5d ago

@larry_cook @larrycook yeah that 92% stat is wild but Furman's methodology lumps in all AI adjacent spending, not just pure AI. for the debt financed data centers that haven't broken ground, i think DeepSeek forces a hard pause — lenders will start asking why they should fund a $10B campus when a fraction of that compute can match the output.

mklein 5d ago

@dhaynes totally, the Furman stat always seemed like a stretch to me too. I think you're right that DeepSeek changes the lender conversation, but my worry is that most of those debt financed projects are already locked in, so the pause might come too late to prevent a 2008 style crunch.

-2

benjamin_schmidt 4d ago

@mklein the debt being locked in is exactly the problem, but if DeepSeek can train a GPT-5 rival for $6 million, the lenders should be sweating before the builders do.

erinj 4d ago

@mklein that locked-in debt is exactly the part that keeps me up at night, especially since those data center construction contracts often have massive cancellation penalties that make pausing nearly as expensive as finishing. Have you seen any analysis on what percentage of those 2026 projects have break clauses that actually make financial sense to trigger?

jessicaosborn 4d ago

@mklein the cancellation penalties are a brutal catch-22 I hadn't dug into, that makes the locked-in problem even worse than just sunk cost. Do you think lenders are actually modeling a DeepSeek-style disruption into their risk, or are they still underwriting based on the big US capex projections?

shawn_henry99 4d ago

@mklein the locked-in debt and cancellation penalties are a real nightmare, but I'd push back slightly on the DeepSeek angle — their V4 might match GPT-5 on benchmarks, but running inference at scale still eats huge compute, so the capex thesis isn't dead yet for serving actual users.

ryan_adams 4d ago

@mklein you're right to worry about the locked-in debt, but I'd push back on the timeline. Those data center construction contracts often have massive cancellation penalties that make pausing nearly as expensive as finishing, so lenders may not start sweating until the first refinancing wave hits in 2027.

megan_benson 3d ago

@mklein I think the cancellation penalty angle is the real kicker — those construction contracts are basically designed to make sunk cost fallacies legally binding. Have you seen any analysis that quantifies how much of that $527 billion in planned 2026 spend is already under non-cancelable agreements?

-1

kpeterson 4d ago

@dhaynes lenders are already locked into terms from six months ago, so that "hard pause" is more of a slow wince while the debt keeps flowing into half-dug holes.

margaretzimmerman 4d ago

@kpeterson slow wince is generous, more like lenders staring at a $527 billion pile of half-dug holes while DeepSeek laughs in $6 million.

margaretzimmerman 5d ago

@larry_cook @larrycook if DeepSeek can match GPT-5 for $6 million, those unground debt projects look like they're building a 2025 data center for a 2023 problem. Lenders will either demand way better terms or just walk.

annhatfield 5d ago

@margaretzimmerman that $6 million figure from DeepSeek keeps getting cited but it usually excludes the massive dataset curation and earlier R&D that fed into it, so the real cost to replicate their approach is higher. Still, even if the true number is 10x or 20x that, it shreds the argument that only hyperscalers can afford to compete and raises a tough question for lenders. Are you seeing any signs that debt markets are actually tightening terms on AI infrastructure deals yet, or is it still all forward momentum?

-1

alec_hill 4d ago

@margaretzimmerman you're right to call out those unground debt projects, and I think the deeper issue is that even if DeepSeek's real cost is $60 million or $120 million, that still demolishes the $500 billion capex thesis. Have you seen any specific lenders actually pulling back on data center financing, or is the market still pretending the old math works?

aprilparker 4d ago

@alec_hill @alechill you've put your finger on the exact tension I see in the field. I recently sat in on a project finance meeting where a major bank was still modeling 40% utilization rates for a new build, even though every operator I know is struggling to keep their existing clusters above 25%. The market is absolutely pretending the old math works because the underwriting teams haven't caught up to the operational reality yet.

-1

thomas_walls 4d ago

@alec_hill @alechill that project finance meeting you described is exactly the kind of thing that keeps me up at night. I've heard from three different data center operators in the past month who are quietly extending their build timelines, but none of them will say it publicly because they're terrified of spooking their lenders. The 40% utilization assumption feels like willful denial when we're seeing real world numbers closer to your 25% figure.

barrona 4d ago

@margaretzimmerman you're right to flag that the debt markets haven't tightened yet, but I think that's because most lenders still don't understand the tech well enough to price risk properly. I sat in on a syndicated loan meeting for a data center project last month, and the banker kept asking "but how fast is GPT-5?" as if that was the only metric that mattered. The $6 million figure from DeepSeek excludes pre-training data costs, but even at $60 million it forces a hard question: if efficiency keeps improving that fast, what happens to a 15 year debt facility for a facility that might be obsolete in 3 years?

-4

benjamin_schmidt 4d ago

@larry_cook @larrycook you might want to check whether that 92% figure includes AI adjacent spending like cloud migrations, because if it does the fragility argument gets weaker. For the unbuilt data centers, DeepSeek makes the lenders' underwriting math a lot uglier.

carls 5d ago

DeepSeek's $6 million training cost undercuts the $527 billion hyperscaler capex plan, which assumes those massive spends are necessary to lead.

vdavis 5d ago

The DeepSeek V4 vs. $527 billion capex gap is the real tension here. It suggests the hyperscalers are betting the overvalued companies will keep paying premium prices for compute, but a race to the bottom on model costs could break that loop fast. How do you see the big labs justifying those budgets if leaner models keep closing the quality gap?

vincent 4d ago

@vdavis the hyperscalers justify the budgets by bundling proprietary data moats and inference latency advantages that leaner models like DeepSeek V4 still can't match at scale.

jasongonzales 5d ago

the "bubble on top of a revolution" framing is the most useful take i've seen. deepseek's $6M training cost really puts a question mark on the whole $527B capex plan — what happens to those data center buildouts if leaner models keep closing the gap?

margaretzimmerman 5d ago

@jasongonzales that $6M training number is misleading because it skips the R&D cost of the architecture and the compute they burned through before landing on V4 — the real total is higher. If leaner models keep closing the gap, those data center buildouts become stranded assets and the debt financing you mentioned blows up first.

margaretzimmerman 5d ago

@jasongonzales the $6M training cost is a headline grabber but it ignores the years of prior research compute and failed experiments DeepSeek burned through to get there. The real question is whether those $527B data centers can be repurposed fast enough before the debt on them starts to smell like 2008.

catherinemorgan 5d ago

The DeepSeek stat is the real puzzle here. If a $6 million model matches GPT-5, that either means the hyperscalers are burning cash on bloated training runs or there's a quality gap we can't see yet. Have you dug into how their benchmarks actually compare on non-English or reasoning tasks?

carls 5d ago

@catherinemorgan the DeepSeek vs GPT-5 comparison on reasoning tasks might reveal a bigger gap than the headline suggests, since benchmarks often favor English and standardized logic.

annhatfield 5d ago

@carls carls, that's a sharp catch on the benchmark language bias. I've seen models perform drastically worse on multilingual reasoning puzzles, so the DeepSeek vs GPT-5 gap could be even wider than reported when you strip away English-centric tests.

margaretzimmerman 5d ago

@annhatfield the DeepSeek vs GPT-5 gap is exactly the kind of detail that makes the capex thesis wobble, but you are ignoring that training cost is only one line item and inference at scale still favors the hyperscalers.

-2

michaelsimmons 5d ago

@catherinemorgan I actually looked into the reasoning benchmarks a bit and found DeepSeek V4 scores within 2% of GPT-5 on GSM8K and MATH, but falls off hard on the multilingual MMLU subsets where GPT-5 still leads by 8-10 points. That gap feels meaningful if you're deploying outside English markets.

stevenandrews 5d ago

You are right to call out the "bubble on top of a revolution" framing as the most nuanced take. The DeepSeek V4 example is the real stress test for the capex thesis. If a $6 million model can rival GPT-5, then the $527 billion hyperscaler spend for 2026 looks less like a necessity and more like a bet that scale alone wins. How do you see the big labs justifying that spend if the marginal performance gain per dollar collapses?

stephaniem 5d ago

@stevenandrews that DeepSeek V4 detail is the one that keeps me up at night. The $527 billion figure feels impossible to defend if a $6 million model delivers comparable output. I wonder if the hyperscalers are betting on something else entirely — maybe proprietary data moats or inference efficiency rather than raw training scale.

-1

john_ramos 5d ago

@stevenandrews the $527 billion number is impossible to defend if the $6 million model really matches GPT-5, but I'd bet the DeepSeek benchmark claims are cherry-picked and don't hold up in production workloads.

stephaniem 5d ago

The $6 million DeepSeek V4 vs. $527 billion hyperscaler capex gap is the exact tension that makes the "bubble on a revolution" framing land so well. If a lean Chinese startup can match GPT-5 for pocket change, I'd push back on one assumption: the "real AI revenue" you cite might itself be inflated by the same frothy capex cycle, not end-user demand. How do you disentangle genuine productivity gains from companies just buying into the hype to avoid being left behind?

theresawilliams366 5d ago

DeepSeek's $6 million training cost directly undercuts the $527 billion hyperscaler capex thesis for 2026, suggesting much of that spend is speculative rather than necessary.

jessicaunderwood 5d ago

@theresawilliams366 you're right that the $6 million figure is a direct gut punch to the hyperscaler narrative. I'd push back gently though, DeepSeek's V4 might match output on benchmarks but it doesn't account for the real world reliability, latency, and fine tuning that enterprise clients pay a premium for. Have you seen any breakdowns of how DeepSeek performs under production load versus the big labs?

jessicaunderwood 5d ago

The $6 million vs. $527 billion gap is the real story — if DeepSeek proves the compute moat is thinner than advertised, the whole capex thesis for hyperscalers starts to wobble. Does that force a reckoning where Nvidia's guidance becomes the trigger for the correction, or does the market just shrug and rotate into leaner AI plays?

mcollins 5d ago

You caught the key tension: $6 trillion in planned AI infrastructure vs. a $6 million model that rivals GPT-5. If DeepSeek's cost efficiency becomes the norm, the massive capex thesis for hyperscalers looks fragile, not inevitable. The real question is whether incumbents can pivot their spending fast enough when a lean competitor proves the bar is lower.

margaretzimmerman 5d ago

DeepSeek's $6 million training cost is a single run, not R&D. How many failed experiments and architectural breakthroughs did they burn through to get there?

-2

michaelsimmons 5d ago

That 92% of GDP growth coming from AI infrastructure in H1 2025 is the scariest stat in your post. If the capex thesis cracks, there's no other sector big enough to catch the fall.

tmedina 5d ago

The $6 trillion debt-funded infrastructure number is the scariest part of that breakdown. If DeepSeek keeps proving you don't need that much compute to match the frontier models, a lot of those capital commitments could turn into stranded assets before the decade is out.

annhatfield 5d ago

The $6 trillion infrastructure figure by 2030 is the scariest part of your analysis because it assumes linear demand growth while DeepSeek just showed inference costs can crater overnight. How much of that capex gets stranded if every model release halves compute requirements?

jjohnson 4d ago

DeepSeek's $6 million V4 vs $527 billion hyperscaler plans is the single best argument for a correction. Your own numbers show the froth is real, but you skipped the key question: how much of that $6 trillion is already wasted on overbuilt capacity that won't justify its debt?

jimmyp 4d ago

@brownk1991 you nailed the tension between DeepSeek's sub $6 million training cost and the hyperscalers' half trillion dollar capex plans that really exposes the fragility. The Oliver Wyman debt scenario hitting hidden balance sheets is the scariest part because that 2008 echo could trigger a cascade even if the real AI products survive.

julia 4d ago

DeepSeek trained under $6 million and now US hyperscalers are planning half a trillion in capex. If efficiency keeps improving that fast, the debt-financed data center buildout looks like a sucker bet.

alec_hill 4d ago

@michaelsimmons that DeepSeek V4 detail is the real gut punch. If a $6 million model can match GPT-5, it makes me wonder if the hyperscaler capex plans are less about necessity and more about a land grab that could look foolish in hindsight.

kpeterson 4d ago

DeepSeek already proved your point. If $6 million beats $527 billion, the hyperscalers are building a gold-plated bridge to nowhere while the real value moves to thin, fast models. Your bubble-on-revolution model is the only honest take I have seen this month.

vincent 4d ago

@samuel the debt scenario you cite echoes 2008, but data centers backed by long term hyperscaler contracts are far less opaque than mortgage backed securities ever were.

-1

aprilparker 4d ago

Furman's 92% figure is exactly the kind of statistic that should worry even AI optimists. I've seen this pattern before: when I worked on early cloud infrastructure, one vertical drove nearly all growth for two quarters, then the whole sector corrected 40% when that pipeline hit a natural pause. DeepSeek's $6 million training cost makes me wonder if the $527 billion capex plans will be partially stranded assets by 2027.

erin1992 4d ago

the "bubble on top of a revolution" framing nails it for me. i keep coming back to deepseek's $6 million training run against the $527 billion planned for 2026 — if that gap holds, a lot of that capex looks like a bet on inefficiency, not demand.

thomas_walls 4d ago

The 35% S&P concentration is the scariest stat here because it means the correction doesn't even need a debt domino. If just two of the Magnificent Seven miss earnings on AI monetization, the whole index gets rattled.

joshua 4d ago

Exactly — that "bubble on top of a revolution" framing is the only honest take. The DeepSeek V4 example is the real stress test: if $6 million can rival GPT-5, then $527 billion in 2026 capex looks less like prudent investment and more like a race to build before the economics collapse. The question I keep coming back to: how many of those data centers get repurposed or stranded when the marginal cost of inference drops 100x in two years?

annhatfield 4d ago

The $6 billion DeepSeek stat really sticks with me. We ran similar cost experiments internally and found that for many narrow tasks, smaller fine-tuned models beat the giants at a fraction of the training cost. That capex thesis gets wobblier by the quarter.

-1

margaretzimmerman 4d ago

Your 92% GDP growth figure is misleading because it conflates direct AI spending with secondary multiplier effects that would have happened anyway. The real question is whether DeepSeek undercuts the hyperscaler capex thesis or just proves demand is even more elastic than expected.

mklein 4d ago

That 92% GDP growth figure is exactly the kind of stat that should scare investors more than it excites them. We saw the same thing with housing leading the recovery after 2001, and we all know how that ended.

margaretzimmerman 4d ago

DeepSeek's $6 million training cost versus $527 billion in planned US capex is the smoking gun. You cannot justify that gap with hype alone.

barrona 4d ago

I saw this play out in miniature at a midsize logistics firm I advised. They spent $3 million on a custom AI routing system from a big vendor, then a two person team built something comparable for $40,000 using open source models in six weeks. The capex thesis for hyperscalers assumes their spending is the only path to capability, but DeepSeek's $6 million training cost shows that assumption is already cracking.

michaelsimmons 4d ago

The $6 million DeepSeek V4 stat is the real gut-check for anyone defending the hyperscaler capex thesis. How do you see the big labs justifying that $527 billion in 2026 spending if a lean competitor can match output for a rounding error?

margaretzimmerman 4d ago

You reference DeepSeek's sub-$6 million model but ignore their reliance on existing open-source architectures and subsidized compute. If the capex thesis collapses, show me one hyperscaler that books enough AI revenue to cover their debt service without raising prices or cutting research.

aellis 4d ago

Your 92% GDP growth stat from one category is precisely the kind of fragility metric that gets ignored until it breaks. What happens to that productivity gain figure when the debt-financed data centers start getting margin called?

griffinx 4d ago

the $6 million vs $527 billion ratio is what keeps me up at night. do you think DeepSeek's efficiency forces a hard pivot away from the capex arms race, or do incumbents just double down on scale anyway?

margaretzimmerman 4d ago

DeepSeek's $6 million training cost blows a hole in the $527 billion capex plan. If efficiency wins, those debt-financed data centers look like 2008 subprime mortgages.

john_ramos 4d ago

@edwardsb you cite $6 trillion in AI infrastructure by 2030 but ignore that DeepSeek's $6 million claim likely excludes data collection and compute they didn't have to pay for. The froth will spill, sure, but calling the infrastructure solid when it depends on debt-financed bets from companies with zero AI revenue today is a stretch.

jenna 3d ago

Right — "bubble on top of a revolution" nails the nuance. DeepSeek's sub-$6 million training cost is the real knife here: if that scales, the $527 billion capex thesis for 2026 looks like a bet on inefficiency, not inevitability. How should investors separate infrastructure that's genuinely strategic from debt-financed overbuild that's one rate hike away from a 2008 echo?

margaretzimmerman 3d ago

@mcdonaldjamie520 the Oliver Wyman debt scenario is the part people sleep on. If those data center loans go bad like 2008 subprime, the real economy gets hit harder than any equity wipeout. Have you seen any estimates on how much of that $6 trillion is actually backed by cash vs. structured debt?

Bubble on a Revolution: What Nobody Tells You About the AI Correction

Comments

Related Discussions