The biggest momentum sale in history
- The worst momentum print in 27 years
- Token volume is shifting to cheaper models
- Micron, , and C are racing to add capacity
…and how to benefit from these trends.
There is a significant bifurcation between fundamentals and the momentum index, creating an outsized opportunity. On one hand, the Morgan Stanley Tech Momentum Index registered a 17-day rate of change of -35.9%, the worst in its 27-year history; on the other hand, TSMC posted record revenue and raised its capex budget yet again.
Prices and fundamentals are telling two different stories right now. And which one you believe determines whether this is one of the best entry points this cycle.
A Factor Unwind Is Not a Fundamental Break
When everything sells off at once, the useful question isn’t “what went wrong?” It’s “who is selling, and do they know something?”
This week’s sellers were overwhelmingly quantitative funds, and what they “know” is only what their models tell them: momentum rolled over, so reduce exposure.
After a first half in which the momentum factor returned roughly 57% and positioning reached the 100th percentile of the past five years, the trade had become extraordinarily crowded, and crowded trades don’t need bad news to unwind.
A quarter-end rebalance and thin summer liquidity were enough to send every fund running the same playbook toward the same exit, leaving Goldman’s High-Beta Momentum basket down 24% month-to-date, its worst run since April 2009.
A model reducing exposure tells you nothing about order books. So we checked the things that would actually make us reconsider and none of them moved:
- Hyperscalers are still guiding to ~$1.4 trillion of capex by 2028. Morgan Stanley just raised its ’27–’28 estimates.
- Micron’s entire HBM4 supply for 2026 is already sold out on long-term contracts.
- is growing revenue at an exceptional pace while locking up virtually every available piece of AI infrastructure it can access… even as cheaper models are gaining share!
Mechanical selling creates prices that fundamentals never asked for. That gap is where fundamental investors get paid.
Cheaper Models Are a Demand Story
The bear case feeding the unwind runs roughly: Chinese models now cost a fraction of frontier pricing:
- DeepSeek V4 Flash runs under $1 per million tokens
- versus ~$50 for Anthropic’s Fable 5
…so AI revenue collapses and the buildout stalls. We’d flip the logic.
In every major technology cycle, collapsing unit prices are what mass adoption looks like, and the evidence says this cycle is no exception.
Low-cost alternatives have been on the market for months; over that same stretch, Anthropic grew to a revenue run-rate exceeding $100B exiting this year, kept raising capital, and hyperscalers accelerated their buildouts.
The workloads with the highest stakes: coding agents, scientific research, financial analysis, healthcare, can’t trade capability for price, so the premium tier keeps its pricing power even as the budget tier expands the market.
Think of it like airline seats.
On OpenRouter, top-10 open-source models fly roughly 89% of the tokens but collect only ~24% of the implied revenue; compared to two frontier models carry ~11% of tokens and ~76% of the dollars.
And this implied revenue split is likely understated as a large share of frontier usage goes direct to OpenAI/ Anthropic APIs and not counted in this estimate. So, real-world frontier revenue concentration is likely even higher.
Who generates the tokens is not the same as who generates the revenue.
Economy fills the cabin; first class pays for the flight.
But every seat needs an aircraft: whether a token is generated for $50 or $0.90 per million, it runs on GPUs, memory, and networking. That’s why we expect hardware demand to triple over the next two years, with memory, CPUs, and networking growing even faster.
The Capacity Race Is On
If demand were breaking, the suppliers would be the first to blink. Instead, they’re all adding capacity:
- TSMC (TSM): reported record Q2 revenue of $40.2B on Thursday and raised 2026 capex to $60–64B, at least $4B above prior guidance, plus an additional $100B investment in Arizona. Full-year revenue growth guidance moved up to 40%+ from 30%+. Chairman C.C. Wei: “Our conviction in the multi-year AI megatrend remains very high.” The stock still fell ~4% premarket as investors took profits after a ~40% YTD run — the momentum paradox in miniature.
- SK Hynix: just completed the biggest international NASDAQ listing ever, raising $26.5B, a record for a first-time U.S. share sale by a non-U.S. company; shares opened ~14% above pricing. The proceeds feed directly into AI capacity expansion.
- Micron (): with HBM4 sold out through 2026, the question isn’t demand; it’s how fast new wafers can come online.
The bigger shift underneath: the agentic AI era is broadening memory demand well beyond HBM.
AI agents need working memory and context the way humans do, which is lifting the entire DRAM complex and data centers are moving from a 1:12 or 1:16 CPU-to-GPU ratio toward 1:1, a CPU opportunity we think could reach $200B versus original projections of $30–45B.
Every new wafer of capacity requires billions of dollars of semiconductor manufacturing equipment, which is where we see some of the best risk/reward right now.
This is what makes the semi-cap equipment companies particularly well positioned to benefit.
Put the week’s three stories together and the picture resolves: the quant complex sold the theme at the very moment the supply chain doubled down on it.
Dislocations like this reward investors who know which businesses are actually getting stronger, and this week, that list got longer, not shorter.