Underclassic   rss   archive   buyer's guide

Issue 85: SpaceX Borrows $40 Billion to Buy AI Chips for 2027

Permanent copy of issue 85. Today's issue is always the front page.

Previous issue  |  This week  |  Next issue  |  archive

SpaceX has closed $40 billion of debt to buy graphics processors for artificial intelligence, Elon Musk said on Sunday, with the chips due in the first half of 2027. In Seoul, Bernstein estimates that Samsung sold $8.7 billion of high-bandwidth memory last quarter, overtaking SK hynix, while South Korea's energy minister said the government will cut speculative data centre demand from its national power plan. On the local side, Nanbeige released a 33-billion-parameter coding model with open weights under the Apache 2.0 licence.

Date: 2026-10-26
Updated: 2026-10-26 06:30
Status: LIVE


Broadcast

The money for central computing is now raised years ahead of delivery. SpaceX has borrowed $40 billion for chips that arrive in 2027, and Samsung and SK hynix are selling every stack of AI memory they can make. Yet the first official doubts are appearing: South Korea's energy minister says most of the data centre demand in his country's plan has not been firmed up. On the desk, a recruitment company's research lab has given away a capable coding model that anyone may run on their own hardware.

Watch


Stories

SpaceX closes $40 billion of debt to buy AI chips, Musk says

Elon Musk said on X on Sunday that SpaceX had closed $40 billion of debt financing, mostly callable debt, to buy graphics processors that are due to arrive in the first and second quarters of 2027. Musk also said debt markets have largely closed to companies without substantial revenue outside AI data centres, and predicted that most of more than 200 smaller AI cloud providers will fail or be absorbed.
-- A single company borrowing $40 billion for chips that will not arrive until next year shows how far ahead the largest buyers are reserving supply. If Musk is right that credit is drying up for smaller cloud providers, the computing that remains will sit with fewer and larger owners.
2026-10-25 / CNBC / chip / central / #spacex-closes-40-billion-of-debt-to-buy-ai-chips

Samsung may have overtaken SK hynix in AI memory sales, Bernstein estimates

Bernstein estimates that Samsung Electronics earned $8.7 billion from high-bandwidth memory in the July to September quarter, up 33 percent, against $7.6 billion for SK hynix, up 11 percent. The estimate rests on South Korean export data, which showed HBM exports up 131 percent from a year earlier in September; both companies report results this week.
-- High-bandwidth memory is the stacked memory that sits beside every data centre AI chip, and exports more than doubling in a year shows where the spending is going. The same factories make the ordinary memory in laptops and desktops, whose prices have been rising as capacity shifts to AI parts.
2026-10-26 / The Korea Herald / chip / central / #samsung-may-have-overtaken-sk-hynix-in-ai-memory

South Korea will cut speculative data centre demand from its power plan

Climate, Energy and Environment Minister Kim Sung-hwan told the National Assembly on Monday that the 12th long-term electricity plan will count only confirmed AI data centre demand, after the ministry nearly tripled its 2040 estimate to 11.9 gigawatts in August. Of 522 data centre projects seeking 3.4 gigawatts that applied for grid studies, only 10, totalling 1 gigawatt, have secured power.
-- Power plans built on announced projects commit households to paying for plants and lines that may never be used. A government discounting its own data centre forecast is a sign that the central buildout is running ahead of what companies have actually committed to build.
2026-10-26 / Seoul Economic Daily / energy / mixed / #south-korea-will-cut-speculative-data-centre-dem

Nanbeige releases an open 33-billion-parameter coding model under Apache 2.0

Nanbeige LLM Lab, part of the Chinese recruitment platform BOSS Zhipin, has published Nanbeige4.3-33B and its base model on Hugging Face under the Apache 2.0 licence, with a context window of 524,288 tokens. The lab says the dense model leads 17 of 20 benchmarks against open rivals of similar size; the scores are self-reported and some competitor figures in its table differ from those labs' own cards.
-- Open weights under a permissive licence let any company run a capable coding assistant on its own machines without paying per token. At full precision it needs about 66 gigabytes, so it reaches a single desktop only through 4-bit versions, which the community has begun to publish.
2026-10-25 / Nanbeige LLM Lab (Hugging Face) / efficiency / local / #nanbeige-releases-an-open-33-billion-parameter-c


Axis
local [#########################-----] central
Score: 82   clusters still winning   unchanged
SpaceX's $40 billion of chip debt and Samsung's surge in AI memory sales show capital and components still flowing to the largest buyers, and Musk himself expects most smaller cloud providers to fail as lenders pull back. Korea's decision to count only confirmed data centre demand is a first official discount on the central buildout, and Nanbeige's open model adds to the desk, so the score holds at 82.

Forces

ForceMeterScorePullNote
CHIP / CAPEX
..
97 central SpaceX says it has closed $40 billion of mostly callable debt for GPUs arriving in 2027, and Bernstein estimates Samsung's third-quarter HBM revenue at $8.7 billion against $7.6 billion for SK hynix.
ENERGY
..
97 central South Korea's energy minister says the 12th electricity plan will count only confirmed AI data centre demand, after 522 projects seeking 3.4 gigawatts applied for grid studies and only 10, totalling 1 gigawatt, secured supply.
EFFICIENCY
..
72 mixed Nanbeige released Nanbeige4.3-33B, a dense 33-billion-parameter coding model under Apache 2.0; its full-precision weights need about 66 gigabytes, so desk use depends on 4-bit versions.

Line charts (0=local 100=central 2026-10-09 to 2026-10-26)

Axis score line chart

Forces line chart

Market overlay (2026-09-27 to 2026-10-26)

Latest: BTC $80,350 / hashrate ~1081.96 EH/s / token markers avg $5.2583/M / axis 82. These lines are different units. We compare timing, not dollars-to-hashrate.

Market overlay line chart

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 $80,350 / hashrate ~1082 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 2026-10-22; blended = (3*input + 1*output) / 4; -- = open/self-host or no single public list)

#ModelOrgCCWeightsTierInOutBlend
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

Full 41-model board


If useful, support the broadcast.
bc1qfs3yw5qlq8sxs50crzh3g84ug27gvww9npu85n
BTC QR

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