Replying to @AtlasShrug1
Coreweave has long term lease at 24 & 25 prices. Not market time priced
Replying to @2whazzuup
Everything that happened in the data center world today (Oct 20, 2026)
Record orders, sold-out chips, and the stocks fell anyway.
NEOCLOUDS + FINANCING
- HSBC rates CoreWeave Reduce, $39 target, 36% below Monday's close. It sees an 8.9% average return on capital for 2026-30, "below its cost of borrowing." HSBC's CEO: "You have to look at the quality of the tenancy agreements" (CNBC)
- JPMorgan says banks' CoreWeave risk is mostly secured: ~$24B of project loans ($15B drawn, $9B undrawn) paid out on milestones. The $21B of unsecured parent debt is taking the hit (The Market Reporter)
- IREN's co-CEO says its data centers carry no debt. Prepayments on recent 3-year deals cover 45-55% of GPU cost, and 2027 capacity is several times oversubscribed (Stocktwits)
- Korean AI cloud firm Elice fell ~30% on debut despite 388x institutional demand. Firmus, AirTrunk and SB Energy shelved IPOs this month (CNBC)
POWER + GRID
- Google will pay for ~$600M of Entergy Arkansas grid and battery work (1.4GWh) plus a 440MW solar PPA, after regulators rejected Entergy's own deal for that solar farm. Needs state approval (Entergy, DCD)
- Google Cloud's London region went down overnight after a "power event." Revolut, Lloyds and UK court video hearings had problems. Still active at 6:22 AM PT (BBC, Google)
- A Texas study finds on-site gas engines emit about 4x the NOx of grid power for the same data center. Large Texas projects could need an estimated 439GW from the grid (Texas Tribune)
- Wisconsin's settlement: full environmental reviews for new data centers and a statewide study by Jan 2028, paid for by Vantage (Wisconsin Examiner)
NEW PROJECTS
- Amazon plans €16B+ in Germany in 2027, including €4.8B to expand AWS Frankfurt (Amazon)
CHIPS
- ByteDance, Alibaba and Tencent ordered 3M+ RTX PRO 5500 cards, $35B+ at list. Nvidia told partners they may get just 5-10%, and it can make no more than ~3M worldwide in 2027 (Reuters)
- Modal says Nvidia's Vera CPU hit up to 2x the throughput of a comparable Intel x86 server on agent sandboxes under heavy load. It's adding Vera, keeping x86 (Modal)
MEMORY
- Micron's Taoyuan union plans to strike Thursday over bonuses. The site makes DRAM and HBM. No length given (Reuters)
- Q4 DRAM contract prices seen +20%, after ~80% in Q2 and ~40% in Q3, as LTA price caps kick in. If Rubin Ultra went to 8-high HBM, the same DRAM could in theory feed 50% more GPUs (Commercial Times via TrendForce)
PARTS + PACKAGING
- Unimicron's chair: some key tool lead times hit 30 months, and low-CTE glass cloth, CCL and ABF could be ~20% short in 2027. CoWoS capacity up 6.8x from 2024 to 2028, substrates only 2x (Wealth Magazine via TrendForce)
- 3M gave Amazon a warrant for up to 3.16M shares at $164.89. 85% vests as Amazon buys optical and other data center products, fully at $6B spent (3M filing)
EARNINGS + MARKETS
- ABB orders rose 40% to a record $12B, data center orders more than doubled, guidance went up. Stock fell ~6% (Reuters)
- KOSPI -3.73%. SK hynix -4.23% to its lowest close since April, Samsung -2.98%. Foreigners net bought Samsung, net sold SK hynix (JoongAng Daily, Yonhap)
What it all means: ABB posts a record and drops ~6%. Elice is 388x oversubscribed and falls 30% on day one. Nobody's arguing about demand anymore. The fight is over who's good for the money. The best credits win twice. Google pays its way to the front of the grid line, and 3M hands Amazon a warrant just to win its business. Everyone else gets sorted by who their tenant is. The number I'd watch is CoreWeave's $9B of undrawn bank loans. Drawn on schedule, the panic stays in the bonds. Milestones slip, that's the real crack. And DRAM going +80%, +40%, +20% while workers want a bigger cut starts to look like a price top, even with supply tight. Lam tomorrow tells us if tool orders care.
$CRWV $IREN $NVDA $MU $SKHY $SSNLF $GOOGL $AMZN $MMM $ABBNY $ETR $LRCX
Follow @SingularityRes for the full data center roundup every morning.
Adobe has better potential than Meta in AI 🤮🤮🤮
“.. At the moment, we have more compute capacity than we have demand, and we have a lot of the tokens and processing power being given away in freemium solutions,” he told @CNBC. “Everyone in the AI race is running for a last-man standing position.”
cnbc.com/2026/10/19/gartner-…
Everything that happened in the data center world today (Oct 19, 2026)
Chips bounced. The bills for power, fiber and debt kept going up.
NEOCLOUDS + FINANCING
- Follow-up: Microsoft confirmed it left Nscale's Loughton site, calling it a "single site" decision. It's still "committed to our partnership with Nscale" (AFP)
- Same day, Nscale is in talks on a ~$45B Anthropic deal for a second West Virginia building, 460MW of Vera Rubin. Terms could change (Bloomberg, Reuters)
- Some investors want Nscale's IPO at $16-19B vs the ~$30B it's seeking, citing Firmus' pulled IPO. Filing: $103.4B of contracts, $140.6M H1 revenue, $1.02B net loss (Bloomberg, Quartz)
- Goldman upgraded CoreWeave to Buy, target cut to $106 from $139. Its case: most debt is asset-backed project loans, so blended debt costs rise only 100-200bp. Stock still closed -0.4% (CNBC, Reuters)
- Oracle 5-year CDS is 295bp, up 287% in a year, a level more like BB/B credit. S&P rates it BBB-. Consensus FCF: -$23.7B this year, -$36B in 2027 (CNBC)
POWER + GRID
- Amazon and AEP will explore up to ~960MW of X-energy SMRs in Indiana, 12 units of 80MW. Amazon funds it and buys first (DCD)
- Israel capped new data centers at 200MW after 27GW of hookup requests. Developers must post NIS 2M (~$643K) per MW requested (Calcalist, DCD)
- Google's 780MW Minnesota deal with Xcel: Google pays for at least 75% of contracted capacity even if it uses less, and owes up to 8 years of those charges if it exits early. Needs state approval (CNBC)
- A Cardiff developer wants a "temporary" 49.9MW gas plant, just under the 50MW line that needs Welsh government sign-off, running 2028-2033 until the grid arrives (BBC via DCD)
NEW PROJECTS + PUSHBACK
- Alcoa is selling part of its old Massena, NY smelter to a data center buyer for up to $315M, only $50-65M at close. Permitted for 180MW, up to 635MW later. An NYDIG affiliate filed plans for the site (DCD, Alcoa call)
- A Virginia bill would freeze all data center approvals statewide July 2027 to July 2030 (Virginia LIS, Loudoun Now)
CHIPS
- Google's TPU v9 tape-out may slip to January after design changes, shipments to 1Q28. Brokers see AI ASIC shipments +69% to 14.1M in 2027 (Commercial Times via TrendForce)
MEMORY
- Lenovo, HP and Dell signed memory LTAs. One exec said they got only 70-80% of the DRAM and SSDs they asked for in 1H, and doesn't expect big problems next year (ETNews)
PARTS + PACKAGING
- Samsung Electro-Mechanics signed a ₩360.5B 2027 memory substrate deal with an unnamed customer, its 7th big LTA since May, ~₩4.3T total (company, Korea Herald)
- China's PCB chain is booming: Kinwong guides Q3 profit +205-265% YoY, China Jushi nine-month profit +100-110% (Commercial Times)
NETWORKING + OPTICS
- TrendForce sees a 180M fiber-km shortfall this year. Utilization 92%, inventory down to 1 month from 4, data center fiber at 5x+ its 2025 price. New preform capacity takes 18-24 months (TrendForce)
MARKETS
- The SOX rose 1.6%, snapping an 8-day slide of ~15.7%, its longest since 2008. Micron +3%, Nvidia +1.4% (Reuters, MarketWatch)
What it all means: Monday was a relief bounce. Nothing got fixed. Nscale loses Microsoft at one site, talks 460MW with Anthropic the same day, and buyers still want half its IPO price. Demand isn't the fight anymore. Who pays for the power and the debt is. Look at the new power terms: Google paying for capacity it might not use, Israel charging $643K a MW just to get in line, a gas plant sized at 49.9MW to skip a permit. A grid slot is now something you pay for up front. The shortages keep sliding down the parts list too, fiber at 5x, substrates on LTAs. Pricing power sits with whoever sells the sold-out stuff, not the balance sheets building on top. Big Tech's next capex guides are the next test.
$NVDA $CRWV $ORCL $AMZN $AEP $GOOGL $XEL $AA $MU $SKHY $SSNLF $GLW
Follow @SingularityRes for the full data center roundup every morning.
Everything that happened in the data center world today (Oct 18, 2026)
The chips are still sold out. The money and the megawatts aren't.
NEOCLOUDS + FINANCING
- Microsoft pulled out as anchor tenant of Nscale's Loughton, Essex AI site after grid officials said power isn't expected until the 2030s. Nscale says it's still committed and is looking at other power (Telegraph, Reuters)
- Nscale's Glomfjord site in Norway is about six months late, with 6.4MW of 30MW live (Bloomberg)
- Some of the $38B loan behind Oracle's Texas and Wisconsin builds has traded around 90 cents, some debt tied to its New Mexico site near 80. Insuring $1M of SoftBank debt costs over $40,000 vs under $27,000 at the start of the year. Barclays says risky AI debt spreads are up 50%+ in six months (NYT)
- Junk buyers are increasingly freezing out tech startups after CoreWeave's bonds hit 15% (Bloomberg)
NEW PROJECTS + PUSHBACK
- RegLayer counts 47+ large US data center projects delayed, withdrawn or suspended this year. Bernstein thinks only about 35% of planned US builds get finished (FT)
- Digital Realty plans a ~1GW campus in Kokomo, Indiana: 476 acres, full build possible by 2033, no tax breaks per the city (DCD)
- An Italian mayor halted a 576MW application billed as the country's largest. Its region has 13.13GW of grid requests (DCD)
POWER + GRID
- Pennsylvania's governor petitioned to join PPL's rate case, a first in at least 15 years. He wants big data centers in their own customer class, paying their own costs. PPL's advanced pipeline is 31.8GW (PA governor's office, The Inference)
CHIPS + POLICY
- Taiwan plans US parks for the small suppliers behind TSMC, Nvidia and Foxconn. The first, in Texas, could cost up to $4B (Bloomberg via CNA)
- China's Pengcheng Cloud Brain III began trial runs on 10,000P of Huawei Ascend compute, heading to 16,000P (Tonghuashun)
MEMORY
- Samsung will use TSMC 3nm for the base die of Nvidia's custom HBM, per industry sources. Nvidia asked Samsung, SK hynix and Micron all to do it. The paper says that gets up to 30% more bandwidth and 15% less power than standard HBM4E (Seoul Economic Daily)
- Same report cites UBS: 2027 HBM demand up 91% vs supply about 10% short (Seoul Economic Daily)
- CXMT guided nine-month profit of RMB131-132.5B. Backing out the first half, Q3 sales rose 15-20% vs Q2 but profit only about 1-4% (CXMT filing, Yicai)
- Follow-up: Samsung fell 17.8% and SK hynix 24.3% Oct 12-16, worst weeks since 2008. Foreigners net sold ₩11.2T of Kospi stocks, 79% in those two (Korea Herald)
PARTS + PACKAGING
- Kingboard raised thin glass-cloth CCL and prepreg 15%, its eighth hike this year. Panasonic told customers some general CCL goes up 200% from January (Commercial Times)
- A Goldman basket of Asian AI hardware fell 11% last week. Samsung Electro-Mechanics -23%, Unimicron -14% (Bloomberg)
NETWORKING + OPTICS
- Salience Labs raised $125M for optical circuit switches, with AMD and Intel investing. It's at 64 ports and wants 256 (DCD)
MARKETS
- The SOX fell 3.8% Friday, its eighth straight loss and longest streak since 2008, down 16% since Oct 6 (Investing)
What it all means: Microsoft walking from Loughton looks like a power call, not a demand call. But that's how the squeeze hits neoclouds. The tenant can walk, the developer is stuck with the land and the debt. A lease on a site without firm power is a free option for the hyperscaler. Nvidia is also pushing all three memory makers onto a TSMC base die. Great for Nvidia's leverage, and if it sticks, TSMC gets paid whichever memory maker wins. The number I'd watch is CXMT: 15-20% more sales for flat profit is a seller's own data saying DRAM price gains are slowing. SK hynix on Oct 28 says if that's China-only. The capex number doesn't get cut. The delivered megawatts do.
$MSFT $ORCL $SFTBY $CRWV $DLR $PPL $NVDA $TSM $SSNLF $SKHY $MU $AMD $INTC
Follow @SingularityRes for the full data center roundup every morning.
Everything that happened in the data center world today (Oct 17, 2026)
Lenders want more for AI debt. Suppliers want more for everything else.
NEOCLOUDS + FINANCING
- Anthropic is offering Blackstone a 10% coupon, more after fees, on $9B of junior debt in "Project North Star," a $60B package for about 1.5GW of Google TPUs. The $42B senior piece is guaranteed by Broadcom. Its last deal, four months ago, paid 8.5% (FT)
- JPMorgan and Goldman sold Volta's $5B AI loan at 95 cents, a yield past 11%. $3.7B of it buys about 36,000 Nvidia GPUs for a Norway site run with Bitdeer (Bloomberg)
NEW PROJECTS + PUSHBACK
- Follow-up on EverWind's Nova Scotia campus: the tenant is Anthropic, its first data center deal in Canada (Reuters)
- San Francisco's mayor signed a 45-day pause on new data centers while the city studies them (SF Standard)
- The UK rejected the 1.2GW Drax-powered AI Growth Zone bid. Drax now plans a smaller 100MW grid-fed site (DCD)
POWER + GRID
- Venture Global signed frame deals with GE Vernova and Baker Hughes for up to 16 gas power islands of 720MW each, about $17.2B with grid gear if all are ordered. Mostly not binding; 8 units have reserved factory slots (company release, 8-K)
- Wisconsin approved two gas plants, 1,120MW and 324MW, that We Energies plans to buy for $2.2B+. One commissioner called it "gas for AI" (Daily Reporter, Heartland Post)
- USDA's rural chief says data centers could benefit from its record $3.3B Georgia gas plant loan. He has almost $8.6B in requests one month into the fiscal year (Reuters)
CHIPS + POLICY
- Taiwan plans to make unlicensed AI server and advanced chip exports to China a crime, part of its US trade deal. No start date yet (Bloomberg)
- China proposed export limits on 23 AI technologies, including AI chip design at 14nm or below, stacked memory design and interconnect for 10,000+ accelerator clusters. Exporters must register now (MOFCOM)
MEMORY
- Foreigners net sold over ₩91T of SK hynix this year through Oct 15, now more than Samsung (~₩89T). Its ₩40T buyback is winding down; Q3 lands Oct 28 (Herald Business)
PARTS + PACKAGING
- Hana Securities sees more memory substrate price hikes from late this year into early next, with BT substrates getting tighter until Korean makers' new capacity fully lands, likely after 2H27 (Chosun Biz)
- Samsung Electro-Mechanics says the MLCC capacitance needed per GPU is doubling every year (company post via CBC)
- Its roughly ₩6.8T ABF substrate expansion may face tool delivery delays (DigiTimes sources)
NETWORKING + OPTICS
- An Asian broker's OCP notes: 83-85% of co-packaged optics failures trace to the external laser. Past 576 accelerators the goal is all-optical; from 72 to 576, copper and optics coexist (Commercial Times)
- Aixtron Q3 orders about €275M vs €215M expected, driven by tools for AI data center lasers (company release)
EARNINGS + BANK NOTES
- Schneider Electric Q3 revenue €11.46B, up 16.9% organic, with "triple-digit" data center demand growth. 2026 targets raised for the second time (Reuters)
- BofA: forward hyperscaler capex consensus is now $1T, about 3% of US GDP vs 2.2% in the 1870s railroad boom. Token prices are down about 50% since summer (BofA via Investing.com)
What it all means: lenders aren't saying no to AI, they're saying "pay me." Anthropic offering 10% on junior debt even with Broadcom guaranteeing the senior piece, Volta clearing at 95, CoreWeave bonds near 15%. Fine for labs with real revenue, a squeeze for the leveraged middle. Meanwhile the choke points keep sliding down the stack to substrates, capacitors, lasers and turbines, and some of those suppliers are pushing prices up. My read: capex doesn't get cut from the top, it gets cut wherever the cost of money eats the margin first. The early warning sign is falling token prices, not the next capex guide.
$NVDA $AVGO $GOOGL $BX $GEV $BKR $VG $WEC $TSM $SKHY $SSNLF $SBGSY $BTDR
Follow @SingularityRes for the full data center roundup every morning.
Because your contract is such that if a neocloud makes money using your model, they need to pay the open weight model developers
That’s how they make money
Open-weight AI labs make zero sense to me.
You spend hundreds of millions training a model.
Then you give the weights away for free.
Your plan is to make money on inference?
Why would anyone pay you, when 10 other providers can run your exact model cheaper and faster?
Your margin goes to zero the day you hit publish.
Genuinely, someone explain the business model.
What people are ready to do for GPUs
Sam Altman@samaOct 16United States
United StatesConnected via WebAccount-level information, not a live location or per-post device.
Wow the new roadster looks quite amazing, good job @elonmusk
Everything that happened in the data center world today (Oct 16, 2026)
The money got nervous. The demand didn't.
FOUNDRY + TOOLS
- TSMC Q3 revenue up 50.9% YoY to NT$1.49T. Q4 guided to $52.3-54.3B, 65-67% gross margin. 2026 capex held at $60-64B, but 2026-28 spend "significantly higher" than the prior three years (TSMC release + call)
- C.C. Wei on AI accelerator demand: no sign of slowdown, growth "probably will be higher" than the old mid-to-high 50s CAGR (TSMC call transcript)
- Street now sees TSMC 2027 capex at $83-90B. UBS has $105B for 2028. Morgan Stanley's checks say EUV bookings could double in 2027 (TrendForce brokerage roundup)
- ASML says 2027-28 supply won't fully meet demand. Some customers want to book orders for 2029 and even 2030. 2026 sales floor raised to €44B (ASML Q3 via CNBC)
MEMORY
- Samsung and SK hynix both showed HBM4E at SEDEX in Seoul: 12-high, 48GB, up to 16Gbps per pin. SK hynix says 12-high samples went to major customers in June and are in verification (Chosun Biz)
- SK hynix pushing HBF, NAND stacked like HBM, at roughly 10x the capacity and about half the bandwidth. Final spec targeted January 2027 (VLSI.kr, Chiprook)
- At OCP, SK hynix also showed a second-gen 256GB CXL 3.2 module with 2x+ the bandwidth of the first gen (Yonhap)
NEOCLOUDS + FINANCING
- CoreWeave bonds sold in June at 9.675% now yield about 15%. Shares fell 15% from Monday's open through Thursday's close (Reuters)
- CDS on CoreWeave's roughly $50B of debt is near 1,000bp, which implies about a 63% five-year default probability, up from about 30% a year ago (Reuters/LSEG)
- Jensen in Stockholm: CoreWeave's "strength is improving every day" and Nvidia doesn't need to backstop it. Stock bounced 6% Thursday (Bloomberg, Reuters)
- Nebius and Iren sold off alongside it. Dan Ives thinks 5-10% of planned projects never get built (Reuters)
- Fermi put a number on its TensorWave lease: 15 years, 222MW, $6.5B of contracted revenue. It hasn't closed yet; closing is set for Oct 31. AMD's credit support only starts once construction is done and rent begins (Fermi via Yahoo Finance, DCD)
POWER + GRID
- NiSource signed a 20-year deal with Microsoft for about 1,100MW of new generation from 2028. Microsoft pays for all the generation and transmission (NiSource release)
- DOE's Paducah hub in Kentucky: Brookfield builds a 1.8GW AI campus, NextEra builds 2GW of gas plus 2.6GW of batteries. $100B private investment, phased build through 2032 (DOE via Yahoo Finance)
- NERC's first data center reliability standards passed their final ballot. They cover sites of 50MW+ connected at 100kV+. The segment where Google, Meta and AWS vote said no. Next stop is the NERC board, then FERC by Dec 31 (The Duck Curve)
NEW PROJECTS
- EverWind: $10B+ in Nova Scotia, a 420MW campus with a long-term lease to a "leading frontier AI developer." 95MW in late 2027, 325MW in 2028 (EverWind release)
- Google confirmed as the company doing due diligence on a 111-acre site in Roanoke County, VA. No deal yet, about 180 days to decide (WSLS)
- 3 G Investments filed plans for a 3.3M sq ft campus near Augusta, GA (Data Center View)
What it all means: the supply side has never been more confident. TSMC is talking about capex that could nearly double in two years, and ASML customers want tool slots in 2030. The weak spot is the debt. When a neocloud's bonds trade at 15% the same week the foundry says demand might be even higher, the market isn't doubting demand, it's doubting who pays for the build. I'd expect the money to keep moving toward the people selling shovels with pricing power (foundry, litho, HBM) and toward the hyperscaler-backed power deals, while the leveraged middle has to prove itself one financing at a time. Watch whether the Fermi/TensorWave lease actually closes on Oct 31.
$TSM $ASML $NVDA $AMD $CRWV $NBIS $IREN $SKHY $SSNLF $MU $NI $NEE $FRMI
Follow @SingularityRes for the full data center roundup every morning.
The scariest line in IDC's phone report isn't the 22% drop.
Memory costs for phone makers are up ~350% in a year. But in Q3 they rose only low double digits vs Q2, after 50–100% jumps in prior quarters (per IDC).
Still rising, just slower. Memory bulls should watch that. $MU
2002: EBay buys PayPal (cofounder by Musk)
2026: Elon Musk rumored to bid for EBay along with $GME
Elon is on an ultimate revenge arc
ASML took in more cash last quarter than it sold.
€11.8B of operating cash flow on €11.7B of sales. A year ago it was €0.6B.
Big driver: long-term contract liabilities (money paid or billed for gear not yet delivered) went from €3.6B to €9.3B in one quarter.
Customers are paying to get in line. $ASML
THE BANKS JUST TOLD YOU WHERE THE AI BUILD GETS SQUEEZED
JPMorgan, Goldman and Wells Fargo all talked about data center financing on their Q3 calls today:
• Goldman's Solomon: recent studies say the US data center build could need "north of $10 trillion" through 2032. Compute demand "significantly exceeds supply" right now
• Same call: who the tenant is changes the financing rate a lot. "There's a big difference" between a project built for a hyperscaler and "some of the neo clouds"
• JPMorgan CFO Barnum: AI exposure is "still quite modest" but not "immune to a major change in sentiment." It's not just construction loans anymore, it's increasingly "working capital revolvers"
• Wells Fargo's Scharf flagged "the significant leverage used to finance the build-out"
Said earlier today the next risk to the AI trade is the cost of money for whoever builds on credit. The people writing the checks just said it out loud. Solomon's version: "it may be" that as financing needs grow, rates get pushed up, "and that will affect the returns and the viability of certain projects."
Same megawatts, different cost of capital. A hyperscaler-leased campus and a neocloud-leased campus aren't the same asset anymore, and the lenders are pricing that in before the market does.
$GS $JPM $WFC $CRWV $NBIS
Didn't expect the best AI supply chain check this week to come from a Fed president.
Chicago Fed's Austan Goolsbee told Semafor that AI inflation is "jumping out of its lane." Prices are going up in spaces "not just components in data centers and the directly adjacent industries," and he's "increasingly nervous" it's plain old overheating demand.
The part that got me is buried in the transcript. He met manufacturers in downstate Illinois. They told him they can't get electricians, and there's "a one-year waiting list for all these different kind of electrical component parts, and that's because of the data centers."
Yesterday I said the action at OCP was everything around the GPU. Power, busbars, connectors. Now a Fed president is hearing it from factory owners in Illinois.
Two ways this cuts.
If you sell electrical gear into this buildout, a one-year waitlist is pricing power, and a rate hike doesn't make that backlog disappear overnight. $ETN $VRT $GEV, and $PWR basically sells the electricians everyone's short of.
If you borrow to build, it's different. Goolsbee says AI investment hasn't been very rate sensitive so far, "but it does seem like they're getting more debt-financed." If the Fed pushes rates up, "the cost of investment would get more expensive." Semafor's own take is that Alphabet and Meta have ad cash cows that can absorb a few extra basis points. Smaller builders on credit may not.
Case in point this morning: $CIFR announced $850M of 0% converts, partly for its Barber Lake build, and the stock got hit hard. Zero coupon isn't free money when you pay for it in equity.
He doesn't vote until 2027. But asked about hiking again this month, he said "if inflation remains persistent, then I think it's clear what the Fed has to do." September CPI lands this week.
The next risk to the AI trade probably isn't demand. It's the cost of money for whoever is building on credit. Own the backlog, be careful with the leverage.
Didn't expect the best AI supply chain check this week to come from a Fed president.
Chicago Fed's Austan Goolsbee told Semafor that AI inflation is "jumping out of its lane." Prices are going up in spaces "not just components in data centers and the directly adjacent industries," and he's "increasingly nervous" it's plain old overheating demand.
The part that got me is buried in the transcript. He met manufacturers in downstate Illinois. They told him they can't get electricians, and there's "a one-year waiting list for all these different kind of electrical component parts, and that's because of the data centers."
Yesterday I said the action at OCP was everything around the GPU. Power, busbars, connectors. Now a Fed president is hearing it from factory owners in Illinois.
Two ways this cuts.
If you sell electrical gear into this buildout, a one-year waitlist is pricing power, and a rate hike doesn't make that backlog disappear overnight. $ETN $VRT $GEV, and $PWR basically sells the electricians everyone's short of.
If you borrow to build, it's different. Goolsbee says AI investment hasn't been very rate sensitive so far, "but it does seem like they're getting more debt-financed." If the Fed pushes rates up, "the cost of investment would get more expensive." Semafor's own take is that Alphabet and Meta have ad cash cows that can absorb a few extra basis points. Smaller builders on credit may not.
Case in point this morning: $CIFR announced $850M of 0% converts, partly for its Barber Lake build, and the stock got hit hard. Zero coupon isn't free money when you pay for it in equity.
He doesn't vote until 2027. But asked about hiking again this month, he said "if inflation remains persistent, then I think it's clear what the Fed has to do." September CPI lands this week.
The next risk to the AI trade probably isn't demand. It's the cost of money for whoever is building on credit. Own the backlog, be careful with the leverage.
OCP day one in San Jose, and the GPU was the least interesting part of it.
The action is everything around the GPU now.
CPUs. Nvidia published a design spec for a Vera CPU rack, built for tool calling, orchestration and sandboxed execution. Same morning, AMD showed Helion, a rack that can ship with zero GPUs. When the GPU company puts out a rack built around CPUs, agents eating CPU isn't a niche call anymore.
Optics. Nvidia has been pushing co-packaged optics hard, and it still put its name on a brand new pluggable with Cisco. USFP is built for 400G per lane, 16 lanes to start with plans for 32, about 4x the bandwidth density of 200G-per-lane modules. TE Connectivity and Amphenol are working on the cages and connectors. Nvidia is clearly planning for pluggables to stick around another generation.
Power. The new MGX 1.2 spec supports racks up to 250 kW and adds a liquid-cooled 5,000-amp busbar. Yes, the busbar needs liquid cooling now. That's why Nvidia, Google and Microsoft are pushing 800V DC, and back in August Nvidia counted more than 80 equipment and infrastructure companies already building to that spec. Visionbay's planned 800V K-1 AI factory uses Vertiv power and thermal tech.
Quick math: 250 kW at 50V is 5,000 amps. At 800V it's about 310. Copper doesn't scale, voltage does.
And Meta is adopting Nvidia's Spectrum-6 102.4T switches for its Vera Rubin build-outs.
This morning I said agents eat CPU. Last week Lumentum said it's sold out into early 2029. OCP is the same story from another angle. The AI capex money is spreading out from the GPU to everything bolted around it.
Names I'd watch from today: $AMD for CPUs, $VRT for power, $APH and $TEL for the connectors nobody models. $NVDA wins either way since it's writing the specs.
APPLIED MATERIALS: AGENTIC AI POD = 2x TO 4x MORE WAFERS
From today's SEMICON West Investor Breakfast slides:
• Gen AI pod: ~600 wafers (GB200 SuperPOD basis)
• Agentic AI pod config A: ~1,400 wafers (~2x) — assumes Nvidia Rubin content
• Agentic AI pod config B: ~2,200 wafers (~4x) — assumes Nvidia Rubin Ultra HBM
• Device mix by silicon area: DRAM 46% → 54% → 67% from Gen to Agentic A to Agentic B
• >85% of agentic AI factory content is leading-edge foundry-logic, DRAM and advanced packaging — AMAT says strong leadership in all three
Same breakfast: those three areas expected to drive ~75% of WFE growth 2025–2030; eBeam portfolio ~$2B in CY26; services to grow >20% this year and exceed $12B by 2030.
Nobody's pricing the wafer math. Agentic isn't just a software story for equipment — it's 2–4x more silicon per pod, and most of that incremental stack is DRAM. If you're long memory into 2027–28, this is the demand-side slide that matches the shortage narrative. The $AMAT bet is owning the process steps around that DRAM + packaging stack, not "AI is good for semis."
$AMAT $MU $SKHY $NVDA $TSM
per Applied Materials Investor Breakfast Oct 13
AEP SIGNS FOR UP TO 1.6 GW OF MAINSPRING LINEAR GENERATORS
• Mainspring Energy will supply AEP with up to 1.6 GW of its linear generators through 2034 (Mainspring PR, today)
• AEP already bought the first 190 MW, with deliveries starting in 2027
• Mainspring calls it the largest procurement of linear generators to date
• Same utility signed for up to 1 GW of Bloom fuel cells back in Nov 2024, which Bloom called the largest fuel cell procurement in the world at the time
So the same utility that signed what was then the biggest fuel cell deal ever is now signing the biggest linear generator deal to date. Together that's up to 2.6 GW of power that doesn't come from a big turbine plant.
Quick math: Mainspring's unit is 0.25 MW (250 kW). 1.6 GW is about 6,400 machines. The first 190 MW alone is ~760.
The catch: per Electroneconomics, Mainspring's whole fleet has 150+ years of core run time. That's roughly 69 days of one 100 MW campus. AEP is making a big bet on a short track record.
Utilities used to build one giant plant and wait years. Now they're buying power in 250 kW bricks because they need capacity fast. And Bloom isn't AEP's only option anymore.
$AEP $BE

























