# AI capex — X 热门讨论 (2026-09-25 15:07 UTC)
## @zerohedge (zerohedge) · 09-25 14:47 · ♥45 ↻8 💬9 Anthropic signed an $11.6bn, seven-year deal for CPU capacity on Akamai’s distributed cloud infrastructure, with the potential to expand by another $9bn, taking the total opportunity to around $20bn. In layman’s terms Akamai is renting Anthropic a huge amount of computing capacity — not GPU chips used to train frontier models, but CPU-based cloud infrastructure spread across Akamai’s global network. The agreement is roughly six times larger than Anthropic’s May contract with Akamai and helps explain the renewed interest in CPU, edge and distributed cloud names after the Muse/Instinct-led agentic AI rally.
It does come with heavy capital intensity, including around $5.5bn of capex tied to the deal and roughly $1.7bn of additional 2026 spend to secure key supply-chain components such as memory, but the read-through is clear: frontier model demand is spreading across the infrastructure stack, and Akamai is now being treated as a more credible AI-capacity beneficiary: GS https://x.com/zerohedge/status/2103496638798999608
## @Niubi0824172693 (马牛逼Trader) · 09-24 21:01 · ♥33 ↻7 💬5 半导体这么强，为什么 AVGO 还在磨盘？ https://x.com/Niubi0824172693/status/2103228351531921767
## @Abbycadabby87 (Abby) · 09-25 14:46 · ♥42 ↻3 💬0 I think the biggest gap in how the market is thinking right now is here.
Everyone is chasing AI software, models, and applications. I’d rather spend my time on the underlying physical infrastructure.
Global AI CapEx in 2026 is already at the trillion-dollar level. Compute infrastructure and space infrastructure may look completely different, but that’s exactly why I’ve been tracking $NBIS, $CRWV, and $RKLB for the long term.
Take space infrastructure as the most straightforward example.
The global commercial satellite industry is already capable of producing thousands of satellites a year. What’s missing is enough launch capacity to put them into orbit.
Launch activity at major North American spaceports has increased sharply over the past few years, with some seeing growth of nearly 467%. NASA has also warned that existing launch pads could approach their operating limits by 2028–2029.
That’s the core of the $RKLB thesis.
From an industry perspective, what’s scarce is stable, continuous launch infrastructure.
That’s why even with semiconductors moving up and down over the past few days, space infrastructure hasn’t really been affected.
The difference comes down to scarcity. https://x.com/Abbycadabby87/status/2103496297625821532
## @onechancefreedm (EndGame Macro) · 09-25 10:06 · ♥32 ↻2 💬5 A number of the individual mechanisms here are plausible, but the thesis links them together as though the sequence is almost automatic. The Fed’s balance sheet is indeed expanding again and Treasury holdings are rising, but the increase has been concentrated in short dated bills tied to reserve management rather than conventional QE aimed at suppressing long yields or broadly stimulating demand. So the direction of the balance sheet is clearly higher, but I would still distinguish that from saying the Fed has already returned to full scale monetization.
The bank handoff is also less clear cut than it sounds. Bank loans and leases have risen from roughly $13.0 trillion to about $14.0 trillion over the past year, but Treasury and agency holdings increased by only around $170 billion over the same period. A meaningful share of the lending growth has also gone to nonbank financial intermediaries. That can support markets, but it can also mean leverage is migrating deeper into the financial system rather than cleanly financing households, factories and productive investment.
The bull steepener argument is certainly possible, but a bull steepener is not automatically a bullish credit event. It often happens because the economy is weakening and the Fed is cutting into softer loan demand and rising credit losses. A steeper curve can help bank margins, but banks still care about collateral quality, defaults, capital and whether strong borrowers actually want more debt. Falling short rates also do not guarantee a weaker dollar and immediate liquidity boom. In a global dollar shortage, the Fed can ease while the dollar still rises because demand for dollars overwhelms the rate effect.
Oil is where I think the sequence becomes most fragile. Act One effectively needs a near term Iran agreement to reopen the Strait and push crude materially lower. Yet Trump has already suggested some form of Iran agreement was close or imminent dozens of times this year, including at least 38 instances by June, without a durable settlement following. More recently he has also indicated a deal could come after the midterms, while Iran’s reopening proposal remains conditional on reduced U.S. military pressure and an end to the blockade. A deal is absolutely possible, but I would not assign it enough certainty to make it the keystone of the macro path.
Act Two also assumes the payoff before it is proven. AI capex is spending happening today. Productivity is the return everyone hopes arrives later. AI can ultimately be transformative and still produce a painful capital cycle first if financing stays expensive, power costs remain high, utilization disappoints or the benefits accrue more to users than to the companies financing the infrastructure.
My own sequence is almost the inverse. Energy stays restrictive longer than expected, the Fed remains tight while underlying demand weakens, refinancing and private credit pressure build, something eventually breaks in credit or funding markets, policymakers respond first with targeted liquidity, recession and disinflation become undeniable, and only then do aggressive cuts pull the front end and eventually the long end materially lower. An Iran agreement would accelerate that transition, but I would not bet on it. > 引用 @BittelJulien: Jeff Gundlach laid out the Fed's dilemma this week:
Hike, and the interest bill on all that short-dated debt balloons. Cut, and inflation reignites.
He’s right. But I think it’s the wrong framing, and I haven’t seen anyone unpack this properly yet in response.
Here’s my take…
The US now spends roughly $1trn a year servicing its debt. And with a deficit bigger than the entire interest bill, every dollar of that interest is effectively borrowed.
They’re using a new credit card to pay the interest on the old one. Oldest trick in the book.
Every cycle the principal gets bigger, the refinancing wall gets bigger, and the liquidity it takes to refinance the debt gets bigger.
Once you understand that, the “dilemma” dissolves. There’s only one exit, and it runs through balance sheets.
The Fed is already back at it. It has added around $365bn of Treasuries since December, and the line is still climbing. Call it bill buying, call it reserve management... it's the Fed monetizing government paper.
But the Fed doesn't want to finance this alone. The real plan is to hand the baton over to the banks.
That's what the leverage rule changes in April were for: free up bank balance sheets to absorb Treasuries and, more importantly, to lend. And they are.
Bank loans are up almost $1trn in a year. When banks lend or buy government debt, they expand the money supply. And unlike QE, far more of it reaches the real economy.
Here's the catch though, and it's the whole game...
Banks borrow short and lend long. A flat yield curve does nothing for them. It squeezes the margin on every new loan.
So the Fed is still pulling liquidity higher as a bridge, waiting for the one thing that makes the handoff work: a steeper yield curve. And it has to be the right kind of steep.
What they need is a bull steepener. Front end falling faster than the long end.
Right now we have the opposite problem.
Markets saw Warsh's hike coming. Yields are up across the curve, with the 10-year and 30-year hitting their highest since 2007, but the front end has sold off hardest, flattening the curve. Exactly what banks don’t want/need.
Warsh delivered last week and signaled more to come. But strip oil out and inflation looks a lot tamer. Core CPI is at 2.4% and still edging lower.
This hike was about independence and credibility with the bond market, not broad-based inflation.
Which brings us back to oil…
Trump wants a deal, and he's saying so openly. Iran has put a road map on the table: a phased reopening of the Strait in exchange for the blockade coming off.
And with the midterms less than six weeks away, nobody in Washington wants voters staring at gas prices the way they are right now.
We've seen one deal fall apart already this year, so I'm not taking it on faith... but the incentives have never been more aligned. If the Strait reopens and crude heads lower, headline inflation loses its biggest tailwind and inflation expectations cool.
That’s the pressure valve.
Warsh has shown the bond market he’s serious. Take oil out of the picture and he has room to stop hiking, then reverse course.
The front end rips, the curve bull steepens, and banks finally have the spread to put those freed-up balance sheets to work.
Then the dominoes fall…
A bull steepener pulls the dollar lower. A weaker dollar lets gold run. And when rates, the dollar and oil are all falling together, that's liquidity rising.
Here's why:
Every one of those forces the world to hedge.
A strong dollar forces anyone with dollar debt or dollar assets to pay up to protect against it.
High short rates make it expensive to hedge dollar exposure, which is why foreign buyers like Japan have largely stepped away from Treasuries.
Expensive oil forces airlines, shippers and importers to lock up capital in margin just to hedge their fuel bill.
When all three ease, that hedging demand falls away and the capital sitting behind it gets released.
And released capital doesn't sit still. It gets levered, lent and financialized.
But that's only act one...
The bigger play is Greenspan, mid-90s. The consensus said above-trend growth had to be inflationary. The consensus was wrong.
Greenspan saw what technology was doing to productivity and refused to fight an inflation wave that wasn't coming.
Real GDP ran at 4-5% for years. Core CPI held around 2-2.5%.
He eased, held his nerve through the boom, and only leaned against it late in the decade.
The Nasdaq 100 rose more than 500% from 1996 to 1999.
Warsh has made it clear he believes the same thing. Growth without inflation, because productivity lowers the cost of everything it touches.
Except this time around the productivity engine is AI and robotics, and it will dwarf what the internet did.
That's how you actually escape the debt trap. Not by paying it down. By growing nominal GDP faster than the debt itself. Debt to GDP stops rising, then eventually starts to fall, without a single dollar being paid back.
So does the party end when the need for debasement fades?
I don't think so. I think it changes shape...
In the 90s there was no QE. The Fed’s balance sheet grew mainly to keep up with the economy’s demand for cash.
Instead, the liquidity came from the private sector: bank lending, bond markets and a booming IPO market funding the buildout.
That's exactly where the banks come back in. Today they’re absorbing government debt so the refinancing gets done. Tomorrow they're lending into the AI capex boom. And that boom is only just getting started.
The big four hyperscalers alone are on track to spend more than 2% of US GDP on capex this year, most of it AI. NASA at the height of Apollo peaked at 0.7%. The Manhattan Project at 0.4%. And it's companies footing the bill, not governments, increasingly with borrowed money.
If you’ve followed my work for a while, you’ve heard me say this before, and I’ll keep saying it:
We’ve spent the last few years teaching AI to think. The next decade is about teaching it to move, see and build.
Robots, factories, power plants, grids… and almost none of that hardware exists yet. Someone has to finance it. That’s the banks’ next job.
So what does this all mean?
Risk assets stop rising on a dollar losing purchasing power (debasement) and start rising on an economy that’s worth more (productivity).
Now, act one hinges on the chart below…
WTI has spent the whole year coiling inside this large range. It just tested the top of it near $107 and got rejected.
So long as crude stays below that downtrend, and especially below $110, act one is on track.
If it breaks out and clears $110, it probably means the Strait deal isn't happening. Inflation stays sticky and Warsh loses his cover.
That delays act one. It doesn't cancel act two. The debt still needs rolling and the productivity wave is still coming.
That’s the playbook as I see it right now.
Act one: a Strait deal lands before the midterms, oil moves lower, the curve bull steepens ahead of the Fed, Warsh pauses then reverses, the dollar falls, gold runs, liquidity rises.
Act two: the productivity boom takes the wheel and the banks finance it.
Watch the yield curve, the dollar and gold for confirmation, then own what outruns debasement now and compounds with productivity later: tech and crypto.
The regime changes. The trade doesn’t. https://x.com/onechancefreedm/status/2103425937035481543
## @MithunSarkari (Mithun Sarkar) · 09-25 10:01 · ♥31 ↻4 💬2 ESDS : Q1 Concall Key points :
1/ The story is NOT the Q1 numbers.💡
Revenue 133.7 Cr (+7.3% YoY) and PAT 29.3 Cr (+~14% YoY)👉 is past
QoQ decline looks sharp, but Q4 had a large one-off technical design revenue from the Sharon AI factory.💡
Core business continued to grow YoY.
The bigger story is what happens when AI infrastructure starts scaling.
2/ ESDS is moving from “Cloud + Managed Services” → AI Infrastructure
Business mix is roughly: • IaaS ~51% • Managed Services ~31% • SaaS ~18%
The new growth engine is GPU-as-a-Service / AI factories. Mgmt said international GPU pipeline is 50,000+ GPUs, while domestic order book is ~3,000 Cr.
Domestic biz is currently growing at ~30–40% CAGR
3/ The biggest near-term trigger: 👉 Sharon AI
ESDS has an AI-factory deal involving ~8,200 NVIDIA B300 GPUs in Sydney.
The deployment delay has pushed revenue commencement into Q3 FY27, with mgmt indicating go-live around November.💡
Importantly:👉
The contract is locked for 7 years.💡
Mgmt said pricing is already discovered/fixed for the 7-year period.❗
This can materially change the revenue scale once the GPUs go live.
4/ Why GPU economics could be powerful ?
The key bottleneck today is NOT demand.
It is GPU availability + capacity.
Mgmt said GPU delivery can take 4–9 months.
International customers are currently preferred because they can offer:
• Better pricing • 12–18 months advance payments • Better margins
This can reduce the working-capital burden of a highly capex-intensive AI infrastructure biz.
5/ Capex : ESDS is entering a massive infrastructure investment cycle. Funding sources 👉 Customer advances + IPO proceeds + debt/lending.
The company is also targeting around 1,500 owned GPUs, with deployment expected around Q4 FY27.
Mgmt wants to secure GPUs ahead of potential price increases.
But remember:
More GPUs = more revenue potential AND significantly more capital intensity.
Execution will therefore matter enormously.
6/ New products can improve the quality of growth
ESDS is not only buying GPUs.
It is building its own software layer around the AI/cloud stack.
New/expanded products include:
• Swaraj Cloud – AI/autonomous cloud platform • Swaraj Garuda – Application Performance Monitoring • Swarajayu – Database Activity Monitoring
The strategic objective is important:
Own more of the stack → move beyond commodity infrastructure → create recurring software/service revenue.💡
7/ Margins: this is where the story gets interesting💡
Historically, SaaS has been more profitable. GPU/IaaS margins are currently evolving as the business scales. Mgmt expects infrastructure economics to improve with scale, while specific lease/release arrangements can carry lower margins.
Therefore we sud not extrapolate Q1 margins blindly.💡
The key monitorables are:
GPU utilisation + pricing + funding cost + depreciation + recurring SaaS/managed-service mix.
That will determine the sustainable margin profile.
8/ Mgmt outlook: bullish pipeline, but NO formal FY27 guidance
Instead, it expects growth to come from:
→ Existing domestic order book → 50,000+ GPU international pipeline → Sharon AI factory → ~1,500 owned GPUs → New AI/cloud products → International expansion
Mgmt also expects seasonality to reduce over the next 2–3 years as more AI factories become operational and billing becomes more evenly distributed.
9/ Key risks:
• GPU deployment delays • Very high capex requirements • Customer concentration • GPU price / technology obsolescence • Financing & depreciation burden • International execution • Margin uncertainty during the scaling phase • Sharon renewal risk after the 7-year contract
CRUX : ESDS is transitioning from a conventional cloud/managed-services company into an AI infrastructure + cloud + software platform.
The next 4–6 quarters will tell us whether the GPU pipeline converts into high-return, scalable earnings.
Execution is key.
No Recommendations #ESDS https://x.com/MithunSarkari/status/2103424734704844842
## @holdmybirra (Hernán Jaramillo) · 09-25 12:51 · ♥31 ↻2 💬1 Nunca había sido tan valiosa la capacidad de aprender rápido.
La tecnología está conectando disciplinas que antes parecían mundos distintos.
Empieza estudiando AI y termina entendiendo energía, CapEx, redes eléctricas, infraestructura y hasta los ciclos ferroviarios de hace dos siglos.
La ventaja ya no está solamente en lo que uno sabe.
Está en qué tan rápido puede aprender algo nuevo, conectarlo con lo anterior y cambiar de opinión.
En este mundo, curiosidad + velocidad de aprendizaje es capital.
Ese es el Día D: un día para conectar antes que los demás las ideas que van a definir los próximos años.
https://t.co/X7JBnNWYVX https://x.com/holdmybirra/status/2103467480328859881
## @DeepDishEnjoyer (peepeepoopoo) · 09-25 12:41 · ♥31 ↻0 💬2 it's especially not true while rates are pretty small. nobody is investing in us treasuries for yield to begin with! ai capex just isn't competing for the same type of demand https://x.com/DeepDishEnjoyer/status/2103464788550451516
## @Venu_7_ (Venu) · 09-25 14:40 · ♥32 ↻0 💬0 Semi equipment is a critical layer of the AI buildout - a massive amount of AI capex ultimately flows into the equipment needed to manufacture more advanced chips.
$LRCX $AMAT $KLAC $ASML are finding support around their 200-day MAs.
$ASML remains the strongest of the group. https://t.co/Rt92ygSccg https://x.com/Venu_7_/status/2103494841367970181