Underclassic rss archive buyer's guide
Permanent copy of issue 93. Today's issue is always the front page.
Previous issue | This week | Next issue | archive
Apple said on Monday that Mac revenue rose 24 percent to $10.8 billion in its fiscal fourth quarter, a record year for the Mac, but warned that memory costs, which it called unprecedented, will keep rising and cut its holiday-quarter gross margin to between 45 and 46 percent. The memory squeeze comes from the other end of the market: SpaceXAI is adding two Mississippi sites with about 420,000 Nvidia GB300 chips to its Colossus 2 cluster, and a consultant's study for PJM, the largest American grid operator, puts data centre demand there at 30 gigawatts by 2030 with an 11.5-gigawatt shortfall in supply. OpenAI, for its part, promised that its Georgia campus will give back up to 1,000 megawatts when the grid is strained.
Date: 2026-11-03
Updated: 2026-11-03 06:30
Status: LIVE
Broadcast
Apple's results showed both sides of the same trade. Ordinary buyers bought more Macs this year than ever, many of them to run AI on their own machines, yet Apple now pays more for every gigabyte of memory because the clusters in Mississippi and elsewhere order it by the hundred thousand. The grid tells the same story: PJM's planners expect more data centre demand than they can serve by 2030, and OpenAI's answer is to promise it will step back on the hottest days in exchange for being connected at all. Local computing is growing, but it is buying its parts in a market set by the largest buyers.
Watch
Stories
Apple reported fiscal fourth-quarter revenue of $114.5 billion on Monday, with Mac sales up 24 percent to $10.8 billion and a record year for the Mac. Chief executive John Ternus called memory pricing unprecedented and said costs will rise again in the December quarter, when Apple expects gross margin of 45 to 46 percent, down from 47.9 percent.
-- Demand for computers that run AI at home is at a record, but the households buying them are now paying for memory that the largest data centres bid up first.
2026-11-02 / CNBC / chip, efficiency / mixed / #apple-warns-memory-costs-will-keep-rising-as-mac
Elon Musk said SpaceXAI is expanding Colossus 2 onto two more sites in Southaven, Mississippi, a former police training ground and a former Spam warehouse, which together will hold about 420,000 Nvidia GB300 chips. He said the first will run by the end of November and the second in December, bringing Colossus 2 to about 1.2 million Blackwell chips by year-end.
-- One company adding 420,000 top-end accelerators in two months is the kind of order that sets the price of memory and chips for everyone else.
2026-11-02 / DatacenterDynamics / chip / central / #spacexai-adds-two-mississippi-sites-with-420-000
A forecast by Charles River Associates, presented to PJM's load analysis subcommittee on Monday, projects data centre demand on the grid at 30 gigawatts by 2030 and 82 gigawatts by 2047, concentrated in the Dominion and AEP zones. It finds that without faster additions of supply, PJM could fall about 11.5 gigawatts short of serving that load at its normal reliability standard in 2030.
-- The grid serving 65 million Americans is being told it cannot build fast enough for the data centres already queued, which puts upward pressure on every customer's power bill.
2026-11-02 / PJM Interconnection / energy / central / #pjm-study-puts-data-centre-load-at-30-gigawatts-
OpenAI launched what it calls the Grid Protectors Program on Monday, under which its future campuses will agree with utilities to reduce electricity use during peaks and emergencies. The pilot is a planned campus in Effingham County, Georgia, where it says Georgia Power can call on up to 1,000 megawatts, roughly the demand of 700,000 homes, and where OpenAI will pay the full cost of serving the site.
-- Offering to switch off on hot days is how the largest buyers now win grid connections that local opposition and state cost laws would otherwise slow.
2026-11-02 / The Verge / energy / central / #openai-pledges-to-cut-up-to-1-000-megawatts-at-i
Deep Principle has published the weights of Hydrogen-1.0, a model for chemistry, materials science and life sciences built on Z.ai's GLM-5.3, under the MIT licence. The company reports that it beats OpenAI's GPT-6-Astra on one research benchmark, but its own tested setup runs on eight data-centre GPUs.
-- Open weights let any laboratory inspect and run the model without an API contract, though at this size the laboratory still needs a server rack rather than a desk.
2026-11-02 / MarkTechPost / efficiency / mixed / #deep-principle-releases-open-weights-for-hydroge
Axis
local [##########################----] central
Score: 85
clusters still winning
0
People are buying more Macs than ever to run software at home, but the memory inside them is getting dearer because the largest AI clusters are buying it first, and grid planners now expect demand they cannot fully serve. The score holds at 85.
Forces
| Force | Meter | Score | Pull | Note | ||
| CHIP / CAPEX |
|
98 | central | SpaceXAI says two new Mississippi sites will add about 420,000 Nvidia GB300 chips to Colossus 2, toward 1.2 million Blackwell chips by year-end, while Apple says the resulting memory shortage will constrain Mac and iPad supply and raised its purchase commitments by $33 billion. | ||
| ENERGY |
|
98 | central | A Charles River Associates forecast presented to PJM puts the grid's data centre load at 30 gigawatts by 2030 and 82 gigawatts by 2047, with an 11.5-gigawatt gap in 2030 unless supply is added faster; OpenAI offers up to 1,000 megawatts of curtailment in Georgia. | ||
| EFFICIENCY |
|
80 | mixed | Apple sold more Macs in fiscal 2026 than in any previous year, $37.9 billion in revenue, with its M6 and A20 Pro chips built around a dual 16-core neural engine for on-device AI; Deep Principle's open 753-billion-parameter Hydrogen-1.0 still needs an eight-GPU server. |
Line charts (0=local 100=central 2026-10-19 to 2026-11-03)
Market overlay (2026-10-05 to 2026-11-03)
Latest: BTC $77,012 / hashrate ~1028.94 EH/s / token markers avg $5.2583/M / axis 85. These lines are different units. We compare timing, not dollars-to-hashrate.
Not shown: GPU cloud $/hour, hub electricity, miner hashprice, or lab cost per token -- none of those are a clean public feed.
Power and list prices
BTC $77,012 / hashrate ~1029 EH/s. List-price gap (not a subsidy %): 123.1x list gap (blended) Mercury 3 -> Claude Fable 5.1 ($0.16/M to $20.00/M). (mempool.space + Coinbase/CoinGecko)
Marker prices (USD per 1M tokens; verified 2026-10-31; blended = (3*input + 1*output) / 4; -- = open/self-host or no single public list)
| # | Model | Org | CC | Weights | Tier | In | Out | Blend |
| 1 | Claude Sonnet 5 | Anthropic | US | closed undisclosed |
workhorse | $2.00 | $10.00 | $4.00 |
| 2 | GPT-5.5 | OpenAI | US | closed undisclosed |
flagship | $5.00 | $30.00 | $11.25 |
| 3 | DeepSeek V4.1 Flash | DeepSeek | CN | open ~284B total / ~13B active (MoE) |
commodity | $0.30 | $1.20 | $0.53 |
If useful, support the broadcast.
bc1qfs3yw5qlq8sxs50crzh3g84ug27gvww9npu85n
Subscribe
$ echo you@email.com | underclassic --subscribe
One email per issue. No tracking pixels, no filler.
home | rss | page 2 | page 3 | archive | models | buyer's guide | subscribe | data | board | mcp | llms.txt