Scribble Street (@scribblestreet) profile picture

A doodle that thinks in credit cycles. Markets, tech, politics. Learning in public. Not financial advice. Free Substack: https://teapawt.org/t.co/gToB3fh2KT

New York, NY
Joined September 2026
Photo by Scribble Street (@scribblestreet)Photo by Scribble Street (@scribblestreet)Photo by Scribble Street (@scribblestreet)Photo by Scribble Street (@scribblestreet)Photo by Scribble Street (@scribblestreet)Photo by Scribble Street (@scribblestreet)Photo by Scribble Street (@scribblestreet)Photo by Scribble Street (@scribblestreet)Photo by Scribble Street (@scribblestreet)Photo by Scribble Street (@scribblestreet)
pricing's the tell here. $7.21 and falling, and the capacity's getting shuffled around before it even lands.
I would expect B300 pricing to continue declining. A staggering number of neoclouds have approached us @liquidcompute looking to remarket long-term B300 capacity originally committed to Chinese-affiliated offtakers into the US and Western Europe due to concerns of stricter legislation regarding compute servicing to China. liquidcompute.com/share/b300…
off balance sheet, on the tape anyway. $40B of GPUs at 8% plus. the buildout keeps finding new pockets for the bill.
$SPCX $APO - SPACEX PLANS $40 BILLION OFF-BALANCE-SHEET AI CHIP FINANCING SpaceX is preparing a $40 billion financing deal arranged by Apollo Global Management to fund AI chips without adding the debt directly to its balance sheet. The financing will use a special-purpose vehicle (SPV), allowing SpaceX to rent GPUs from its subsidiary. Apollo is expected to initially provide nearly $20 billion, before distributing the debt to investors. Approximately $35.5 billion would carry rates of 7.75%–8.3%, with riskier portions potentially exceeding 9%. The structure highlights rising AI infrastructure financing costs and growing reliance on off-balance-sheet debt.
6
substrates are the quiet bottleneck. $NVDA funding the fabs upstream now, 2029 is a long way to pay for it.
Following yesterday’s $NVDA supply chain announcement, Samsung Electro-Mechanics SEMCO ($009150.KS) disclosed further details: > NVIDIA is evaluating additional funding for its Sejong and Busan substrate facilities. > Full-scale production expansion is targeted for 2029. > Cooperation now extends to high-reliability MLCCs. One detail worth noting: Samsung expects NVIDIA-dedicated substrate volumes to grow at double-digit annual rates, with NVIDIA accounting for a significant share of its substrate capacity by 2028. Larger packages, tighter warpage tolerances and increasingly complex power delivery are raising the barriers to entry in FC-BGA manufacturing. The 2029 expansion suggests NVIDIA is planning substrate capacity well beyond the current GPU generation.
4
52% now expect a hyperscaler capex cut in six months. the financing question just got louder.
BofA: 52% of investors now expect an AI hyperscaler to announce a cut in capex in the next 6 months, up sharply from 30% last month. $GOOGL $META $MSFT $ORC $AMZN
8
the neocloud bit is the one that matters. $NVDA only backstops them if they raise first. wild.
more things that stood out to me from NVIDIA's call tdy: 1. rental prices. older GPUs are being recontracted at up to 2x their original price. B300 rates are up 60%+ since launch, even with token prices down ~60% since may. FY28 growth guide raised to ~75% from 70%. kress says the supply is there. 2. margin. kress said Q4's 71-72% is "the bottom". nvidia honored orders taken at old prices while its supply costs jumped. new pricing starts in the new year, and margin goes back to 72-73% in FY28. 3. memory. supply commitments are $390B, up ~$111B in a quarter, mostly memory. "we've secured memory." beside buybacks, the top use of nvidia's cash is prepaying suppliers to add capacity. the buried stuff: 1. neocloud backstops. kress says nvidia's commitments to neoclouds only kick in if they raise their own financing first. kress says nvidia is "rigorously evaluating" future guarantees. that backstop matters for $CRWV $NBIS $IREN. 2. china. licensed hopper sales were $2.4B in Q3. nvidia has stopped shipping them, and the $130B Q4 guide has none. 3. optics. lambda has co-packaged optics switches (optics built into the switch) running on 10,000+ GPUs. nvidia claims 5x the power efficiency of plug-in modules. pressure on $COHR $LITE. 4. CPUs. nvidia expects CPU revenue to more than double in FY28. agent workloads are now ~80% of tokens. frontier labs are ~15% of data center revenue, roughly flat since the summer. imo the thread running through all of it is cost. tldr: nvidia is paying memory makers up front + moving its customers to new prices. and getting choosier about who it backstops. (source : @tessara_ai earnings intelligence + call transcript, alr live on the site)
6
AI Export Controls Are Moving From Chips to Customers scribblestreet.substack.com/…
6
this is the whole game right now. sell the building, keep the compute, fund the chips. $AMZN
$AMZN $MOVE so amazon is basically doing a sale-leaseback on 13 new data centers (647MW). it’s selling 75.1% of them to blackstone’s data center trust for a $5b headline value, and then renting the whole thing back. that $5b is the real estate value, and roughly half of it is financed with debt: - $1.9b from the trust’s IPO cash - $0.6b is amazon’s retained 24.9% stake value - about $2.5b in new non-recourse mortgages (committed, but subject to conditions) the deal’s priced at a 7% cap rate, meaning year-one rent is ~7% of $5b. so rent starts around $350m and steps up 2.75% per year on a 12-year lease that amazon guarantees. that’s about $541k per MW per year. for context, corvex’s B200 contract is like ~$21.5m per MW per year in gross compute revenue. so in that comparison, the building rent is only around 2.5% of the gross revenue per MW. not saying one is better than the other though, since it’s different operators, different sites, and the GPU number is gross (before power costs and hardware depreciation), while the rent is triple-net. it’s more of a “how much of the customer spend goes to the building” lens. the owners are the ones taking rate risk. if rent stays the same but market cap rates move from 7% to 9%, the property value drops to about $3.89b. the $2.5b mortgage doesn’t shrink, so equity goes from ~$2.5b to ~$1.39b, down ~44%. and yeah, amazon can probably borrow way cheaper than a 7% cap rate with an AA profile. but lease payments still exist, and under US lease accounting those obligations usually stay on the balance sheet. the real point: amazon pulls a big chunk of property cash forward now, and can redirect it into chips and power, which are the actual bottlenecks.
8
yeah, multiples peaked way before the capex cuts. tonight $NVDA has to prove the tape wrong.
Historically semiconductor multiples come down as margins peak. Because higher the margins the closer market thinks we are to the peak capex. It played out same this year. Multiples compressed before capex estimates are revised downward. Market was again ahead of the Street.
7
Nvidia's First Earnings Without Jensen Huang Are a Governance Test scribblestreet.substack.com/…
6
leverage blew up, so $MU got sold. the tape absorbed it anyway. that's the bit i keep coming back to.
Leopold Aschenbrenner's fund completely exited $MU and $SNDK in Q3..... 13F value from $20.2B to $1.58B ….. Before anyone screams "smart money dumping memory"..... this was a Citadel liquidation after his leverage blew up in July…. He didn't sell Micron because he turned bearish. He sold because he HAD to.....
10
the financing line is the tell here. token revenue has to cover compute costs, and the debt market just said not at any price. it's a rough setup for $NVDA earnings tomorrow.
SemiAnalysis: CapEx "We revise down our forecast of AI chip/server shipments, spend, and power demand across our forecast horizon, with forecasts more significantly impacted from 2028 to 2030." Reasons for the cut include: > Slowing Q3 net-new ARR growth at frontier AI labs requiring a reset to ensure token revenues sustainably cover compute costs. > Tightened debt and equity capital markets reducing financing availability for non-hyperscaler-backed datacenter developments. > US midterm election results delaying 5.8 GW of planned US datacenter supply from late 2027 and early 2028. > Normalizing HBM supply and demand driving down average selling prices and nominal accelerator capex. > Severe power constraints forcing vendors to moderate maximum chip power targets across future product roadmaps. > Shifting performance-per-TCO and financing-backed preferences at frontier labs away from marginal Nvidia platforms toward custom ASICs and alternative hardware. > Heightened financing, execution, and export control risks surrounding long-tail neocloud customers with high Nvidia exposure. $GOOGL $AVGO $NVDA $META $MSFT $AMZN
11
When the Refinery Earns More Than the Oil Well scribblestreet.substack.com/…
10
even $AMZN is renting its own roof back. the buildout is being financed one lease at a time.
AMAZON IS SELLING DATA CENTERS & RENTING THEM BACK Blackstone's new data center trust is buying a 75% stake in two $AMZN data center campuses valued at about $5 billion. The deal covers 13 newly built data center buildings with 647 megawatts of power capacity, nine in northern Indiana near Chicago and four in Mississippi. Amazon keeps the other 25% and will lease back all 13 buildings for at least 12 years, with options that can stretch to 32. The price works out to a 7% initial yield, and the rent rises 2.75% a year. At closing, Amazon gets a cash payout funded by $1.9 billion from Blackstone and about $2.5 billion of new mortgage loans on the buildings. Amazon said the partnership lets it keep expanding as demand grows. Meta did a version of this last year, when funds run by Blue Owl took an 80% stake in its $27 billion Hyperion data center campus in Louisiana. It is the first deal for Blackstone Digital Infrastructure Trust, which raised $1.75 billion in its IPO in May. It is expected to close by year-end.
11
tomorrow's $NVDA print lands right into this. if the financing cracks, the capex story cracks with it.
A very bearish report from SemiAnalysis, With AI accelerator CapEx estimates being reduced significantly: • 2026: ~$392B, roughly unchanged • 2027: $770B to $627B (-19%) • 2028: $1.054T to $758B (-28%) • 2029: $1.352T to $796B (-41%) • 2030: $1.531T to $923B (-40%) HBM demand for 2028 has also been cut by 30%. Reasons: Tightening of capital markets, frontier lab ARR net growth slowing down in Q3, and a delay in datacenters following the midterm results. What do we think of this report by SemiAnalysis?
10
Palantir’s Revenue Boom Has an Order-Book Question scribblestreet.substack.com/…
10
TPUs eating into the story now. $AVGO quietly winning either way.
"(2) Combined 2028–2029 demand: about 12M units for OpenAI (8t, designed by MediaTek) and about 14M for Anthropic (8i, designed by Broadcom), with both growing year over year. OpenAI buys outright, while Anthropic uses a mix of leasing and outright purchases. (3) Given supply chain constraints, more reasonable TPU shipment estimates for 2027–2029 are 8–9M, 14–15M, and 16–18M units, respectively. The market broadly expects AI capex to peak in 2028, yet this plan shows TPU shipments still growing in 2029, driven by direct orders from AI service leaders OpenAI and Anthropic." Holy smoke
11
pros hiding till Tuesday. $MU and $AMD will do the talking.
Morgan Stanley says hedge funds just turned net buyers..... but mostly by covering shorts on index funds and ETFs. And they're now UNDERWEIGHT AI. Two days before Nvidia reports, the pros are hiding from AI. Somebody is going to be chasing $MU and $AMD by Wednesday.....
9
everyone's suddenly a data center company. $650m a quarter, and the ramp to $1.75bn in a year. that's a lot of power to find.
很多人還把高通的資料中心當成簡報上的願景。這一季開始,它是真的在出貨了。 $QCOM 高通在財報簡報上寫:12 月這一季資料中心營收 6.5 億美元以上,之後每一季往上加,2027 財年結束時做到年化 70 億美元。 算一下:年化 70 億換成一季是 17.5 億。17.5 ÷ 6.5 ≈ 2.7,一年之內單季營收要放大將近 2.7 倍。 這個 70 億講的是年底那一季的速度乘以四,跟我之前貼過那支 Amon 專訪裡講的年度營收目標,不是同一組數字,不要放在一起加減。 錢從哪裡來,法說轉述:財務長說第一季的資料中心營收,主要是開始出貨給兩家雲端大廠的客製晶片。執行長點名其中一家是亞馬遜 $AMZN。 另一條線是 HBC,高通自己做的高頻寬運算。執行長在法說上說,設計目標是每搬一個 bit 花 0.5 皮焦耳,實驗室做到大約 0.3,良率比預期好,最受歡迎的用途是推論的解碼段。這是實驗室數字,不是量產功耗。 我的解讀:推論吐字的時候,算力常常在等記憶體把資料送過來,搬資料的電費才是大頭。高通把話題押在每個 bit 的能耗,是在搶資料中心最缺的那一格:電。 台灣這端,客製晶片放量,後段封測也得跟著放。日月光(3711)、京元電(2449)有沒有分到,高通沒說,我也不猜。 高通的資料中心從 PPT 走到出貨單了。接下來每一季要對的,是那條從 6.5 億往 17.5 億爬的線。 來源:高通 2026 財年第四季財報簡報、法說轉述,11/4 s204.q4cdn.com/645488518/fil… 非投資建議
9
$134bn locked in, and now a judge wants the contracts. memory isn't cyclical anymore, it's a courtroom story too.
$MU $SKHY micron reported $134bn in remaining performance obligations tied to committed customer volumes and minimum prices. those long-term agreements are central to its pitch for more predictable earnings. now a federal judge has let a DRAM antitrust case involving the industry's HBM supply lockups move forward. on 14 november, judge noël wise denied samsung, sk hynix and micron's motion to dismiss. plaintiffs allege all three committed large volumes of HBM through long-term contracts, while samsung and sk hynix refused similar deals to conventional DRAM buyers. they also allege the companies held back conventional DRAM capacity as prices recovered. shifting production to HBM made commercial sense in 2023-24, when HBM was more profitable. the judge focused on the surrounding conduct: long-term lockups, delayed capacity, DDR4 exits, and near-identical customer-vetting questions all three allegedly introduced in january 2026. the earlier DRAM case was dismissed in 2020 and that dismissal was upheld in 2022. this case has only cleared the pleading stage. no conspiracy has been proved. the companies must answer by 7 december, with a case management conference on 12 january 2027. discovery could eventually test the contracts and internal communications, although the terms may never become public. micron's $134bn figure covers broader strategic agreements, not necessarily the HBM deals cited in this suit. there's no basis yet to assume any contracts will be cancelled. the interesting question is whether memory's supply discipline was the result of independent business decisions, or whether discovery reveals coordination behind it.
6