The main positions are still working. Bitcoin has held its breakout range, AI usage is growing, and the memory shortage still looks more durable than a normal upcycle.
I do not see enough evidence to cut these trades. I do think it is time to start managing concentration.
Several positions in the book depend on the same two things: AI capex staying on schedule and risk appetite holding up. If either one changes, the first move can be fast. Protection is still reasonably priced, so I would rather buy a little now than wait until everyone wants it.
1. Bitcoin: the breakout is about to meet the sellers
Bitcoin is pressing into the supply zone we flagged last week. The area that matters is still $80K to $82.5K, where a large amount of BTC have their on-chain cost basis, formed on the last leg up before the selloff. Price has spent the week grinding just below it, in the high $70Ks.
Those coins do not all need to trade. The data shows where coins last moved, not who holds them. Some of that supply may sit with institutions and long-term holders who have no interest in selling near breakeven. The market only needs to absorb the price-sensitive part. We will know that is happening if bitcoin can spend time inside this zone with high turnover and limited downside.
The flow data is constructive. US spot ETFs have taken in money for eight straight sessions through August 26, about $2.8 billion over the run, and August is already the strongest month of the year at roughly $3.3 billion. On-chain, the seven-day EMA of net realized P&L is positive at around $752 million, with realized profit near $1.1 billion against $354 million of realized loss. There is supply above, but there is spot demand walking into it.
The healthy path from here is fairly boring. Bitcoin trades between roughly $72.5k and $84k for a few weeks. Leverage and volatility cool down. ETF flows stay positive, Coinbase Premium does not roll over, and coins keep changing hands without a large loss of price.
A mild rise in bitcoin dominance would help focus liquidity on this process, but it is not required. A sharp rise caused by altcoins selling off would be a risk-off signal, not a bullish one.
A sustained close above $82.5K would tell us the auction is clearing. On the downside, the level is $76.6K, which sits near the short-term holder cost basis. A daily close below it is manageable on its own. It turns into a real warning if at least two of these deteriorate with it: ETF flows, Coinbase Premium, and the seven-day EMA of net realized P&L.
We remain constructive. There is no need to chase every move inside the range.
2. AI: usage is growing faster than spending
The latest OpenRouter sample showed token volume up 47% month over month, while spending rose only 7%. Volume-weighted pricing fell 28%, driven by both list-price cuts and a mix shift toward cheaper models. Low-cost models accounted for almost all of the monthly increase in token volume.
Demand is not slowing in this sample. More people are using AI, and they are using more of it. But the growth is concentrated at the cheap end, and that is the question for investors: if the marginal token keeps getting cheaper, how much of the usage growth turns into revenue, and how quickly does that revenue cover the capex?
The hardware data is not flashing a warning yet. In Bloomberg’s monthly index for non-hyperscaler capacity, B200 rental pricing fell 1.5% in August, the first sequential decline in that series in more than six months, while H100 pricing rose 1.2%. One month is not a trend, and OpenRouter skews toward developers and startups rather than the full AI market. Still worth tracking. A broader decline in accelerator rental prices would be an early sign that supply is catching up or that buyers are getting price sensitive.
Memory looks better than the rest of the hardware cycle. And SK Hynix has been outperforming its peers in the recent bounce as we flagged. Recent supplier commentary points to long-term agreements covering roughly 50% to 70% of volume, with the exact share varying by supplier, often with sizable prepayments. That gives suppliers better demand visibility and could make earnings less cyclical than in older DRAM cycles. HBM pricing still has room to surprise on the upside.
The chart has not fully caught up with the fundamentals. That is fine. We remain long memory, but additions should come after the technical repair holds rather than after a fast green day.
The nearer risk is not demand itself but how the market reads the same data. If falling token prices and softer rental rates get read as evidence that capex returns are weakening, crowded hardware positions can sell off before any project is actually delayed.
