---
title: "The Compute Watch - No. 1: Rents Up, Memory Barely Moved"

author: "Lida Liberopoulou"

date: "2026-11-02"

license: "CC BY-SA 4.0"

canonical: "https://threadbaire.com/blog/posts/compute-watch-01.html"

description: "October's numbers: AI chip rents rose on both indexes, the newest chips fastest, while the memory price rose 1.6%. The prediction's test hasn't started."

series: compute-watch

series_part: 3
---



***New here? [Read this first](https://threadbaire.com/blog/posts/compute-watch.html): what this watch is, why it exists, and the prediction it checks. The starting numbers are in [No. 0](https://threadbaire.com/blog/posts/compute-watch-00.html).***

AI companies rarely own every chip they use. Many rent them by the hour from cloud companies and from specialist firms that do little else. A growing share of the money that bought these chips was borrowed, and the rent is what pays the loans back. But chips are built to be replaced. A better one arrives every year or two, and the old one earns less. The owners' books, and many of the loans, assume the chips keep earning for years. This watch tracks, once a month, whether they do. The core question: does the newest chip keep earning much more than older ones once something better arrives? 

## This month 

Rents rose in October on both indexes, and the newest chips rose fastest, so the gap between the newest chip and the older one grew. Memory rose much less. That's consistent with a shortage that is tightening, not easing. My prediction, that the newest chip loses its premium once the shortage eases, can't be tested until it does.

## The ladder

| Chip                              | No. 0 (2 Oct) | No. 1 | Change since No. 0 |
| --------------------------------- | ------------- | ----- | ------------------ |
| A100 (2020)                       | $1.58         | $1.56 | −1.3%              |
| H100                              | $2.77         | $2.93 | +5.8%              |
| H200                              | $3.37         | $3.46 | +2.7%              |
| B200                              | $5.86         | $6.50 | +10.9%             |
| B300                              | $6.78         | $7.56 | +11.5%             |
| *H100 at the big cloud companies* | *$7.20*       | *$7.16* | *−0.6%*          |

*Silicon Data indexes, specialist providers unless marked, as shown on 2 November 2026.*

Every chip rose except the oldest, and the two newest, the B200 and B300, rose the most. At the big cloud companies, the H100 barely moved.

## The numbers

| What | Baseline (2 Oct) | No. 1 (2 Nov) | Change |
| --- | --- | --- | --- |
| **The premium:** B200 rent ÷ H100 rent | 2.12× | 2.22× | Wider |
| **Loan terms:** chip-backed borrowing | Nscale, benchmark + 5 points (floating) | Sharon AI, $356m: 9.95% fixed on $206m; $150m starting at 7.25% and rising to 9.95% (1 Oct, not in No. 0) | Not directly comparable: fixed vs floating |
| **Server lifespans** in the owners' accounts | Alphabet 6, Oracle 6, CoreWeave 6, Meta 5.5, Amazon 5–6, Microsoft 2–6 | Unchanged | No new data |
| **The thermometer:** memory (GDDR6) | $19.31 per GB | $19.61 per GB | +1.6% |

**Loans priced around the month.** A loan's interest rate reflects what the lender is betting on. The baseline loan, Nscale's, is a bet on the chips themselves: if the borrower can't pay, the lender takes the chips. Sharon AI's loan works the same way, so the two can be compared. The others are bets on something else: Nebius's on the whole company, North Star's on Anthropic keeping up its lease payments. Their rates say less about what chips are worth.

| Borrower | Date | Type | Terms |
| --- | --- | --- | --- |
| Sharon AI | 1 Oct | $356m for 3½ years, to buy the chips for one customer's contract; if Sharon AI doesn't pay, the lenders (led by Goldman Sachs) take the chips and the contract | $206m at 9.95% fixed; $150m starting at 7.25% and rising to 9.95% (company filing) |
| Volta | 5 Oct | $5bn loan to an AI cloud company with lower-rated credit, offered by JPMorgan to investors; terms not yet final | About 11% yield, benchmark + 6.25–6.5 points (Bloomberg) |
| Nebius | 17 Oct | $3.5bn for one year; if Nebius doesn't pay, the lender has first claim on almost everything it owns | Benchmark + 2.75 points, rising 0.25 every 90 days to a cap of + 3.5 (company filing) |
| Anthropic "North Star" | 26 Oct | $63bn borrowed by a company set up to buy chips and rent them to Anthropic; Anthropic's rent pays the loan back | Final rates not public. Banks hesitated; it closed at higher rates and with more support from Broadcom (WSJ) |

Two details are worth noticing. Sharon AI's press release gave one rate, 9.95% fixed. Its filing with the SEC shows the loan in two parts, and one part starts at a lower 7.25% before rising, so the announcement gave the simpler picture. And Nebius's rate goes up every three months the loan isn't repaid: the longer Nebius keeps the money, the more the lender charges for the risk.

**Memory moved in steps.** GDDR6 is graphics memory, not the HBM inside these chips; it's used here as a public reading of the wider memory shortage. The index is published daily, but in October it moved in four small steps, roughly a week apart, and was flat in between: $19.31 → about $19.36 (6 Oct) → about $19.47 (13 Oct) → about $19.51 (around 20 Oct) → $19.61 (27 Oct). TrendForce's separate GDDR6 spot price was roughly flat in late October (−0.23% in the week to 26 October). Both say memory rose slowly, if at all, while rents on the two newest chips rose about 11%.

**The two indexes.** Ornn's prices, settled 1 November, rose more than Silicon Data's:

| Chip | Last reading | 1 Nov | Change |
| --- | --- | --- | --- |
| H100 | $2.71 (30 Sept) | $3.02 | +11.4% |
| B200 | $8.09 (30 Sept) | $9.44 | +16.7% |
| H200 | $5.23 (5 Oct) | $6.21 | +18.7% |
| A100 | $0.92 (5 Oct) | $1.06 | +15.2% |

Ornn's premium went from 3.0× to 3.1×. The two indexes moved the same way this month on every chip but the oldest: Silicon Data had the A100 down 1.3%, Ornn had it up 15%. Overall they moved slightly further apart: the gap between their premiums grew from 0.87 to 0.91. On Ornn, the B200 rose steeply in mid-October and has slipped slightly from a late-October peak. Silicon Data's readings show no such slip.

## The test

**The shortage isn't easing, so the test hasn't started.** My prediction is about what happens *after* the shortage: once chips are easy to get again, the newest one should lose its premium over older ones. While chips are still scarce, almost any chip finds a renter at a high price, so the premium tells you nothing about the prediction yet.

To decide when the shortage is over, I use the memory price as a gauge. I set the rule before starting: the shortage counts as easing once memory is at least 15% cheaper than in No. 0 ($16.41 per GB or less) for two notes in a row. This month memory is 1.6% *more* expensive. So the premium widening from 2.12× to 2.22× doesn't count for or against me. It's recorded here, and it will matter only once the test starts.

There's a reason the wait could be long. Three things are holding back new supply at the same time:
- **Money:** lenders are charging the specialist rental firms more, or turning them down (CoreWeave's borrowing costs jumped; Firmus couldn't sell its shares).
- **Power:** sites can't get electricity connected (Loughton in England, NextDC's site in Melbourne).
- **Memory:** it's still getting more expensive.

Each one pushes the end of the shortage further away. If it hasn't ended by No. 6, the prediction stays open, and I'll say so.

## What happened in the month

Things that might later show up in the numbers. None of them is claimed as a cause.

**Prices**

- **Buyers said they were short too.** Alphabet said it started renting outside computing capacity at the end of the third quarter because supply was limited, and will use more in the fourth (28 Oct). Microsoft said customer demand "continues to exceed available supply", and that token prices were higher than it expected (28 Oct).
- **One seller rationed instead.** Users of Grok Bot reported reaching their usage limits much faster than usual. Elon Musk said the cost of serving it "was unsustainably high relative to revenue" (1 Nov).

**Lending**

- **A neocloud's credit was repriced.** In the week to 16 October, CoreWeave's shares fell more than 15%, the yield on its 2030 bonds rose about 5 percentage points to near 16%, and the cost of insuring its debt against default climbed to about 12% a year (Reuters).
- **The biggest chip financing nearly stalled.** Anthropic's $63 billion "North Star" financing, repaid from its lease payments for Google-designed chips, was announced on 26 October. The *Wall Street Journal* reported that several banks had considered pulling out, and that it went through at higher rates, with more support from Broadcom. Final rates weren't published.

**Demand**

- **Anthropic filed its IPO prospectus** (26 Oct). The predictions I made before it was filed are scored below.

**Supply**

- **The next generation became rentable, at one cloud.** Nvidia and Oracle said Vera Rubin systems were generally available on Oracle's cloud (27 Oct). No public rental price was given, and Rubin isn't in either index yet.
- **A neocloud's listing was withdrawn.** Firmus, an Australian firm that rents out AI chips, withdrew its stock market listing after weak demand at its asking price (9 Oct).
- **Power held up two sites.** Microsoft withdrew as the main tenant of Nscale's planned site at Loughton, England. Grid power there was reportedly not expected before the 2030s (18 Oct). NextDC cut its guidance after grid works for its Melbourne M3 site were suspended; they won't restart before February 2027 (21 Oct).
- **Some rented capacity is late.** Microsoft said some contracts for chips rented from outside providers were delayed to January–June 2027 (28 Oct).
- **SpaceX rents at spot.** SpaceX said it aims to rent out about half its AI computing capacity, "at effectively the spot-market rate, as opposed to signing these long-term deals" (20 Oct). On 31 October, Musk said its Colossus 2 site would hold about 1.2 million Blackwell chips by the end of the year. If they arrive and some are rented out, the B300 row is where it would show.
- **Memory is being locked up.** SK hynix said long-term agreements with 14 major customers cover more than half its 2027–28 production capacity, typically for five years (28 Oct). The more memory sits in long contracts, the more the spot price measures a leftover slice. That's the same thin-market problem as chip rent.

**The market itself**

- **Futures are listed but not trading.** ICE has listed monthly futures on Ornn's B200 and H100 prices, from December 2026 to December 2027, and publishes daily settlement prices for them. As of 2 November, no contract has traded. CME's futures on Silicon Data's prices have not launched.
- **The CFTC's comment period closed** (20 Oct). Comments came from CME, ICE, both index makers, trading firms, investors and academics. I filed one ([0060](https://www.regulations.gov/comment/CFTC-2026-1850-0060)), using No. 0's figures.

## Predictions scored: Anthropic's prospectus

On 23 October, before the prospectus was public, I predicted four things it wouldn't let a reader trace ([Follow the Money Both Ways](https://threadbaire.com/blog/posts/follow-the-money-both-ways.html)). Each is a pass if the filing shows less than it could. I scored them against the first public filing (S-1, 26 October).

1. **Use of the IPO money stays vague: pass.** The prospectus names "compute and infrastructure" only as a category of spending. It doesn't link the money to any of the signed commitments to Broadcom, Google, Amazon or Microsoft.
2. **Neither Amazon nor Alphabet in the related-party section for compute: pass.** Amazon appears more than 150 times in the prospectus, but not in that section, which is the only one the test counts.
3. **The Broadcom leases can't be rebuilt to the ~$161.2 billion total: pass.** The prospectus gives one total for the leases. The filed 2026 lease blacks out every price, quantity and tranche total, and sets the interest rate in a side letter that isn't filed. It refers to an earlier lease from August 2025, so there are at least two, but nothing says what each is worth.
4. **Cash invested and commitments can't both be stated for all three: pass.** The prospectus states all three commitments: Google at least $114.5 billion, Amazon at least $110 billion, Microsoft at least $44.9 billion. On the other side, it says only that the three "have made significant equity investments", with no figure for Microsoft's. So a reader can't set what each put in against what Anthropic owes it.

Four of four. I said at the time that these were the likelier bets. What held was the bet that the bills wouldn't be connected.

## The strongest counter-argument

No counter-arguments were sent in this month, so this is the strongest one made in public.

**Chips last eight years, not five or six.** Jensen Huang, Nvidia's chief executive, told Yahoo Finance at GTC Berlin on 21 October that "what people thought used to be a four-year life cycle really is probably eight to maybe more", and that "you might be able to underwrite a loan against the asset itself". It's the strongest objection so far, because it names the watch's own variable. Mark Zuckerberg and Elon Musk have made versions of the same point this month: old H100s are worth more now than a year ago.

**My answer.** A chip has three lives: the years it runs, the years someone uses it, and the years its rent covers what it costs to own and power it. The debt and the depreciation rest on the third. Huang's case rests on demand: "there's always companies that can use the compute", "so many offtakers". That's true in a shortage, when almost any chip finds a renter, and the watch's test only starts once the shortage eases. In the same interview he called power "one of the major bottlenecks". Where power is what's short, each new generation does more with the same power, and the old chip loses its slot. It can run for eight years. Its claim on the power may not last that long.

## Reader question

No reader questions arrived this month. If you have one, [send it](https://threadbaire.com/contact.html#compute-watch).

## What I'm watching next month

**New supply.** Whether the Blackwell chips SpaceX says it is adding by the end of the year come online, and whether the B300 rent softens if they do.

**Whether the late-October easing continues.** Ornn's B200 slipped slightly from a peak around 27–28 October, and a few buyers said capacity became easier to find in those days. If it's more than a few sellers clearing excess, December will show it.

**Memory against rent.** Whether rents keep rising much faster than memory. If they do for a second month, that's worth saying plainly, without changing the test.

**ICE volume.** Whether any of the listed futures trade.

**When Rubin enters the indexes.** It's rentable at one cloud now. When an index adds it, I'll note it as a change in contents, not a market move.

**Power as the stated cause.** This month, two sites were held up by power (Loughton, Melbourne), and Nvidia's chief executive called power "one of the major bottlenecks". I'll note each time a company or a filing names power as the reason for a delay, to see whether it becomes the main constraint instead of the chips.

## Contribute

Send me what you pay for compute, a correction, or a question: [send a note](https://threadbaire.com/contact.html#compute-watch). Your own numbers are never published or put into AI tools, and you don't need to give your name. [How contributions work](https://threadbaire.com/blog/posts/compute-watch.html#how-to-contribute).

You can [get each note by email](https://lidaliberopoulou.substack.com) on the day it comes out.

---

*How this is made.* Headline figures come from Silicon Data's public index page, recorded on the first working day of each month, using specialist-provider prices unless marked. Silicon Data's page probably showed the previous working day's values when it was read, so the figures are given "as shown on" the reading date. Ornn's index is the cross-check, and every Ornn figure carries its own date. The two measure related but different things: Silicon Data combines completed deals with posted prices and adjusts them to a standard configuration, while Ornn uses only completed on-demand rentals. That's one reason their levels differ. Changes are my own calculations from the recorded values, not the percentage changes shown on the index pages. The GDDR6 index is in US dollars per GB; it is published daily but moves in roughly weekly steps. Index changes are noted as changes in contents: this month, Silicon Data added an A100 hyperscaler price and Ornn added the RTX PRO 6000; neither is used here. A screenshot of every page is kept for each note. As is common for benchmark providers, both index makers are backed by firms with a stake in the market they measure: Silicon Data's investors include CME's venture arm and the trading firms DRW and Jump; Ornn's include a16z crypto. All of these have been disclosed by the companies themselves.

*Research, not investment advice.*

**Sources:**

- [Silicon Data GPU Rental Price Indices](https://www.silicondata.com/products/silicon-index) and [RAM Index](https://www.silicondata.com/products/ram-index) (2 Nov 2026), quoted under Silicon Data's [Terms of Service](https://portal.silicondata.com/terms-of-service); [methodology](https://www.silicondata.com/documents/silicon-data-index-methodology.pdf);
- [Ornn Compute Price Index](https://data.ornn.com/markets) (settled 1 Nov 2026); [methodology](https://data.ornn.com/methodology);
- [TrendForce DRAMeXchange, GDDR spot price](https://www.dramexchange.com/) (26 Oct 2026);
- [ICE, OCPI B200 Future](https://www.ice.com/products/83050009/Ornn-Compute-Price-Index-OCPI-B200-Future) and [OCPI H100 Future](https://www.ice.com/products/83050006/Ornn-Compute-Price-Index-OCPI-H100-Future) (2 Nov 2026);
- [CME Group STP notice, 28 Sept 2026](https://www.cmegroup.com/notices/stp/2026/09/20260928.html);
- [CFTC comment 0060](https://www.regulations.gov/comment/CFTC-2026-1850-0060);
- [Anthropic, PBC, Form S-1, 26 Oct 2026](https://www.sec.gov/Archives/edgar/data/0002130357/000162828026067168/anthropicpbc-sx1.htm);
- [Nebius Form 6-K, 17 Oct 2026](https://www.sec.gov/Archives/edgar/data/1513845/000110465926117978/tm2628078d1_6k.htm);
- [Bloomberg, 5 Oct 2026 - JPMorgan dangles 11% yield on Volta's $5 billion AI loan sale](https://bloomberg.com/news/articles/2026-10-05/jpmorgan-dangles-11-yield-on-volta-s-5-billion-ai-loan-sale);
- [Nvidia and Oracle, 27 Oct 2026 - Vera Rubin on Oracle Cloud Infrastructure](https://nvidianews.nvidia.com/news/nvidia-and-oracle-expand-nvidia-vera-rubin-to-multi-gigawatt-scale-and-bring-nvidia-vera-cpu-to-oracle-cloud-infrastructure);
- [SK hynix, Q3 2026 results](https://news.skhynix.com/en/q3-2026-business-results/);
- [Silicon Data Series A announcement, 11 Aug 2026 (press release as reprinted by The AI Insider)](https://theaiinsider.tech/2026/08/24/silicon-data-secures-30-5m-series-a-to-build-the-independent-benchmark-layer-for-the-ai-compute-economy/);
- [a16z, CFTC comment 0098](https://www.regulations.gov/comment/CFTC-2026-1850-0098) (discloses its Ornn investment);
- [SharonAI Holdings Form 8-K, 1 Oct 2026](https://www.sec.gov/Archives/edgar/data/0002068385/000149315226045368/form8-k.htm) and [press release](https://www.sec.gov/Archives/edgar/data/0002068385/000149315226045368/ex99-1.htm);
- [Wall Street Journal, 26 Oct 2026 - Broadcom CEO intervened to seal $63 billion Anthropic debt deal](https://www.wsj.com/finance/broadcom-ceo-intervened-to-seal-63-billion-anthropic-debt-deal-c807754f);
- [Reuters, 16 Oct 2026 - Why CoreWeave is in the crosshairs of the AI scare](https://www.reuters.com/legal/transactional/why-coreweave-is-crosshairs-ai-scare-2026-10-16/);
- [Forbes Australia, Oct 2026 - Firmus' $44 billion IPO collapse explained](https://www.forbes.com.au/news/investing/firmus-44-billion-ipo-collapse-explained/);
- [Reuters, 18 Oct 2026 - Microsoft pulls out of UK Nscale data centre deal over power delay](https://www.reuters.com/business/media-telecom/microsoft-pulls-out-uk-data-center-deal-with-nscale-telegraph-reports-2026-10-18/);
- [NEXTDC, M3 Melbourne power delay and revised FY27 guidance, 21 Oct 2026 (AEDT)](https://live.deutsche-boerse.com/news/EQS-Adhoc-NEXTDC-Limited-M3-Melbourne-Power-Delay-and-Revised-FY27-Guidance-7ecf06f9-08b6-4088-a7bf-0a3c547b9c34);
- [SpaceX, Q3 2026 webcast transcript, 20 Oct 2026](https://s21.q4cdn.com/184289198/files/content_files/SpaceX-Third-Quarter-2026-Webcast-Transcript.pdf);
- [Alphabet, Q3 2026 earnings call, 28 Oct 2026](https://abc.xyz/investor/events/event-details/2026/2026-Q3-Earnings-Call-2026-Ujou16CTw9/default.aspx);
- [Microsoft, FY2027 Q1 earnings call, 28 Oct 2026](https://www.microsoft.com/en-us/investor/events/fy-2027/earnings-fy-2027-q1);
- [Elon Musk on X, 1 Nov 2026 - Grok Bot](https://x.com/elonmusk/status/2116992258520064397);
- [Elon Musk on X, 31 Oct 2026 - Colossus 2, 1.2 million Blackwells](https://x.com/elonmusk/status/2116596535433871630);
- [Yahoo Finance, 21 Oct 2026 - Jensen Huang at GTC Berlin](https://www.youtube.com/watch?v=sH5qyeHSdVs) (lifespan remark at about 6:42)