Just read through your article. The bullwhip framing is an interesting lens and I have not seen anyone else apply it here.
The compute-price-as-leading-indicator point is also quite sharp, I just think our market is still very much in market participant and price/liquidity discovery mode. The signal works well though in theory because of the storage thing mentioned above. In a storable commodity, inventory decouples price from demand and a shortage can sit in a warehouse for months before anyone sees it. Compute has no buffer, so a demand shock hits the rental price the same week. That is why your Blackwell move leads the meter rather than lagging it. We see the same pattern one layer down in the physical market, where contract terms move before prices do. Prepay requirements and payment timing tighten first, then price follows.
One place I think is interesting and worth a bit more thought. You put conversion at 35 to 40 percent and give two reasons for the missing 60. One is that demand is real but stuck behind constraints. The other is the rationing game, buyers over-ordering on purpose because they expect to be cut. Those look identical in queue data and mean opposite things. If it is constraints, the load arrives late and PJM was early. If it is padding, it never arrives and PJM was wrong.
While the compute market is young and growing, it has some similar traits actually. Offtake letters of intent are cheap to sign and get counted as demand, and the drop-off between LOI and signed contract is where you find out which is which. That gap is visible to those of us in the market day-to-day before it is visible anywhere upstream, and it is a cleaner read on real demand than anything in the queue data imo.
Would enjoy comparing notes if you keep writing on this, thank you for sharing!
The utilities parallel is a good one. Compute cannot be stored unlike oil, so there is no cash-and-carry anchoring the forward to spot. The curve is expectation plus a risk premium, same as power.
Your infra builder point is super interesting. A data center developer is short power and long compute at the same time, and today they hedge one leg and not the other. Locking power for fifteen years while the revenue side reprices every two or three is a duration mismatch as is….
Downstream has the mirror image of it but that’s where market participants need the most education imo. Buyers (sometimes) hedge the contracts they can see, but the bigger exposure is renewal. A two-year contract against a five-year product commitment means year three is unpriced, and most enterprises are not modeling that as market risk at all.
Right, and worth remembering CDS didn’t start with a DC either. It started with bespoke confirms and bilateral disputes over what counted as a restructuring, and the machinery got built because dealers needed to net and settle consistently. The definitions followed the volume, not the other way around.
The compute version is a bit different though to your point. There’s no event to determine because obsolescence isn’t binary unlike an EoD. Instead, chips gets slowly less useful and then suddenly a lot less useful, and a continuously observed price absorbs that better than a determinations committee vote could.
What we’d actually need a determination process for is disruption, not obsolescence. If the underlying market stops printing or a spec disappears, someone has to decide what settles. Another interesting problem....in the meantime you're right the gap is very tradeable!
Yeah this is a fair statement, they are two distinct things. But they arrive in one quote because the seller only gives you one number, and that’s true in every term market.
Take a 5y swap rate, a forward oil price, or 3y CDS spread…these all contain expectations plus a premium for commitment, and they’re not separable from just looking at them.
You need a second instrument. Once a cash-settled forward trades alongside the physical, the premium becomes observable instead of modeled.
Game 1 was doomed from the start.
If the Dodgers legitimately just get 2 HITS with RISP, they win Game 2.
If they get just ANY production from Mookie and Freddie tonight, they likely win Game 3.
Starting pitching was bad, but LA easily could have taken these last two games.
The Dodgers:
Won 106 games in 2019 and were eliminated at home by the Nationals in the NLDS
Won 106 games in 2021 and were eliminated by an 88-win Braves team
Won 111 games in 2022 and were eliminated in the NLDS by an 89-win Padres team
This might be the most humiliating
Mookie Betts and Freddie Freeman went a combined 0-for-21 this series
We’re not counting that infield single/error
Flat out embarrassing from your two superstars
Inexcusable
UConnPlayed the overall toughest collective opponents between R64 and R32 and won both CONSIDERABLY
Keep in mind DD victory and 70 pts against St Mary’s (slow slow slow pace and elite defense) is high level. Would agree but you need Uconn up there with ease and SDSU in 2nd tier
Not letting Tyler Anderson go another inning
Pulling Yency Almonte mid at bat after 2 massive outs
Evan Phillips nowhere to be found
I just don’t get it