The AI Trade Is Finally Cracking | Weekly Roundup

The yen carry trade unwind and the momentum factor implosion are structurally intertwined and tend to happen simultaneously. The momentum factor saw a four standard deviation move on the unwind day, and the current setup is described as closely mirroring July 2024, when the yen unwind coincided with what is called a Fed hawkish pivot mistake. The Fed stayed hawkish after a cold CPI print before reversing at Jackson Hole and cutting rates. The argument is that the current Fed committee has repeated that mistake by projecting potentially three hikes even as labor is rolling over and inflation is expected to follow. Moving from two hikes priced in to one or none is only marginally better for liquidity and is not equivalent to rate cuts pumping markets. A chart comparing inflation swaps to Fed sentiment via language model analysis shows the Fed turned hawkish right at the pivot point, mirroring the 2024 pattern exactly.

The jobs report was mixed in a concerning way. The unemployment rate fell not because of healthy labor demand but because labor force participation shrank. NFP showed significant misses, wage growth is described as down with no signs of revitalization, and ISM PMI data came in weaker with the new orders versus inventories ratio turning downward. There is described as no fundamental case for the Fed to hike into this labor market. All forward-looking indicators point to inflation rolling over, with oil back below 70 dollars, seasonal dampening, and waning fiscal impulse in the second half identified as underappreciated headwinds. The key risk that could prevent core PCE from rolling over is the wealth effect from AI-related asset prices flowing through into broader inflation.

The Mag Seven had been showing weakness for weeks before cracks spread more broadly. Amazon announcing potential capex cuts caused weakness to spread from large-cap leaders into semis, AI names, and semiconductor markets in Korea and Taiwan. Meta considering selling excess AI compute capacity was enough to raise doubts about the infinite compute demand narrative. An anonymous but reportedly reputable account claimed a significant architecture breakthrough around memory efficiency by a team spun out of OpenAI, and that single tweet was described as sufficient to unwind the entire AI trade. These catalysts only had the impact they did because of extreme momentum positioning already built up in the AI semiconductor trade. Approximately 100 billion dollars of triple-levered ETFs were moving AI-related markets removed from fundamentals, and anyone bearish on the AI trade had already been blown out to the upside. Stocks in Japan, Korea, and Taiwan in the semiconductor space reversed two to three months of price action in days, which is described as characteristic of a bubble dynamic.

The trade is also reflexive in a way that amplifies downside. Approximately 30 percent of hyperscaler income earlier in the year came from markup of AI lab private valuations on their balance sheets, meaning lower AI receiver prices are also negative for the hyperscalers who are the payers. Open questions hanging over the trade include why Meta is renting out excess compute and why OpenAI delayed its IPO. The next two months are described as the most treacherous period going into the back half of the year, with the economic reacceleration trade and the AI bull trade at risk of unwinding simultaneously as liquidity wanes and the Fed remains too hawkish. Record yen shorts, dollar length, and SOFR shorts are described as all colliding simultaneously. The posture expressed is to sell rips rather than buy dips, with a negative bias on Nasdaq and tech.

Gold is favored as a trade expressing the fading of Fed hawkishness as real yields appear to be peaking. The fundamental case for debasement trades in gold and Bitcoin is described as as strong as ever, with US deficits remaining at approximately 6 percent of GDP, debt continuing to climb, and issuance being manipulated. Bitcoin sentiment has been negative for roughly a year, making it susceptible to short squeezes, though reaching above 100,000 would likely require Michael Saylor's participation. The prediction offered is that over the next six months the AI trade will likely fade, which would benefit debasement assets. A caveat is raised that the shift of retail participants from crypto into AI means a crypto resurgence will take longer than past cycles to materialize.

On crypto market structure, most tokens are described as worthless and the industry is seen as needing to wash out misaligned token and equity structures. The Venice token was cited as a specific problematic example where equity holders can sell into token buyers who hold no claim on equity. Hyperliquid was offered as a contrasting case, having built a strong following because its founders were perceived as not exploiting token holders. The old dynamic of all crypto assets rising together has largely ended, meaning value will not necessarily accrue across tokens even if Bitcoin performs well.