Capex Circuit

Tracking the AI buildout through chips, GPU demand, HBM, data-center capex and earnings. Numbers over narratives. No stock calls.

[Capex Circuit @RaphFieldNotes]

What if the AI bottleneck is memory, not GPUs?

70% of high-end memory production capacity could be consumed by AI hyperscalers in 2026. That is the key break from the old memory cycle: prices rose, fabs added capacity, supply overshot, prices crashed.

The constraint is HBM wafer intensity. HBM uses roughly 2-3x more wafer area than standard DRAM, so incremental AI server demand both absorbs scarce HBM and removes potential output from conventional DRAM.

Reported contract terms point to a structurally tighter market than a normal spot squeeze. Samsung, SK Hynix and Micron are pushing 5-year base agreements with rolling extensions, with coverage moving toward 60-70% of planned capacity. Pricing bands include supplier-friendly floors; Samsung reportedly caps quarterly price cuts at 5% while allowing more room for increases. Micron has around $22B in take-or-pay deposits; SanDisk has $11B+ in guarantees.

The pressure is already showing up outside HBM. In Shenzhen’s Huaqiangbei market, DRAM spot quotes rose 10-14% in one week after months of quiet trading. DDR4 8Gb 3200 moved from $19.5 to $22. DDR5 16Gb and 24Gb rose about 14%.

The infrastructure read-through: this is not only an Nvidia demand story. If HBM capacity is effectively sold out through 2026, and UBS is right that DRAM supply-demand may not balance until Q2 2028, the second-order exposure sits in memory allocation, prepayments, substrates, testing, and legacy DRAM supply.

The question for AI infrastructure costs: which line items reprice quietly while the market is still focused on GPU headlines?

[Capex Circuit @RaphFieldNotes]

What changed at AMD in one year?

$6.7B reported data-center revenue, doubled YoY, is less interesting than the margin move: -5% to 31%. Segment operating loss of $200M became $2.1B income—clear AI-infra operating leverage, or peak-cycle arithmetic?

[Capex Circuit @RaphFieldNotes]

What if the AI capex bubble is backwards?

$793B: 2026 hyperscaler capex estimates, up from $546B 8 months ago. MSFT/GOOG/AWS show ~$1.7T cloud backlog. With GPU/HBM + power constrained, risk may be underbuild, not overbuild. Disagree?

[Capex Circuit @RaphFieldNotes]

Can AMD prove the AI GPU boom has a second winner?

Aug. 4 after the close is a market-wide AI infrastructure check, with AMD earnings as the vehicle.

The first line item to watch is data center growth, followed by management’s language on AI GPU demand. Strong guidance would indicate hyperscalers are still willing to fund a second supplier at scale. Soft guidance would carry read-through for Nvidia, HBM, servers, power, and other assets priced around rising AI capex.

The key issue is mix. EPYC server CPU strength can make the quarter look solid while GPU traction remains hard to quantify; Ryzen AI PCs can create narrative momentum without adding much near-term revenue. Shipment visibility, backlog quality, and margin direction matter more than another broad “demand is strong” comment.

The non-consensus framing: AMD does not need to “beat Nvidia” for the print to matter. It needs to show customers want credible alternatives badly enough to buy in volume. That would affect pricing power, supply negotiations, and how investors value the AI hardware chain.

Expect volatility either way. Perpetuals positioning is pointing to the same setup: the headline print may matter less than the first 10 minutes of guidance. What would make AMD’s AI GPU story credible here: revenue, margins, customer names, or guidance?

[Capex Circuit @RaphFieldNotes]

Can AMD prove the AI GPU boom has a second winner?

Aug. 4 after the close is a market-wide AI infrastructure check, with AMD earnings as the vehicle.

The first line item to watch is data center growth, followed by management’s language on AI GPU demand. Strong guidance would indicate hyperscalers are still willing to fund a second supplier at scale. Soft guidance would carry read-through for Nvidia, HBM, servers, power, and other assets priced around rising AI capex.

The key issue is mix. EPYC server CPU strength can make the quarter look solid while GPU traction remains hard to quantify; Ryzen AI PCs can create narrative momentum without adding much near-term revenue. Shipment visibility, backlog quality, and margin direction matter more than another broad “demand is strong” comment.

The non-consensus framing: AMD does not need to “beat Nvidia” for the print to matter. It needs to show customers want credible alternatives badly enough to buy in volume. That would affect pricing power, supply negotiations, and how investors value the AI hardware chain.

Expect volatility either way. Perpetuals positioning is pointing to the same setup: the headline print may matter less than the first 10 minutes of guidance. What would make AMD’s AI GPU story credible here: revenue, margins, customer names, or guidance?

[Capex Circuit @RaphFieldNotes]

Why is AI's optical bottleneck hitting before CPO?

Goldman: optical cable/device & InP substrate names are up 200–350% YTD, rivaling DRAM/NAND. Demand still mostly pluggables/telecom; CPO/NPO laser scale looks 2028. Bottleneck yes, CPO cycle not yet.

[Capex Circuit @RaphFieldNotes]

Can you call AGI and still get margin-called?

165 pages made Leopold Aschenbrenner a 2024 SV must-read: Situational Awareness framed the AI buildout. This week his AI hedge fund got squeezed as chip stocks fell. Durable capex thesis—or sizing too early?

[Capex Circuit @RaphFieldNotes]

Why is Amazon's AI boom showing up in CPUs?

$496B backlog, up 154%: Amazon’s Q2 report reads like an AI infrastructure bill coming due. Revenue was $200.6B, up 20%. AWS was $42.2B, up 37%, marking its fifth straight quarter of acceleration. Amazon also said chip revenue and AI revenue are both above $25B annual run rates, each growing triple digits.

The under-discussed AWS disclosure: customers want inference beside their applications and data. That shifts AI spend into storage, vector databases, networking, security, and core compute. Once a model moves from demo to production, latency and data-movement costs become architecture constraints.

The CPU detail matters. Amazon said post-training RL and agent tool use mostly run on CPUs rather than AI accelerators, and that Graviton delivers 30–40% better price performance. If agents are calling tools, reading databases, rewriting context, and running eval loops all day, the meter is running outside the GPU cluster as well.

The infrastructure read-through: AWS does not need to own the single frontier model to monetize AI demand. Bedrock can route customers across Anthropic, Meta, Mistral, and open models while AWS sells the layer where inference, proprietary data, and application logic meet.

Question for builders: are you optimizing for the model leaderboard, or for the part of the stack that gets expensive after users arrive?

[Capex Circuit @RaphFieldNotes]

After Nvidia, which AI bottleneck gets paid next?

Nvidia was the cleanest AI infrastructure exposure because the bottleneck was clear: larger training runs required more GPUs. The next phase should be less linear. As GPU supply improves, the binding constraints shift to memory bandwidth, advanced packaging, networking, power, cooling, and lower-cost inference.

The better screen is cost per token. Does the company make model serving cheaper, faster, or improve cluster utilization? If not, it is likely AI veneer. A liquid-cooling supplier tied to contracted data-center builds may have a clearer capex read-through than a software vendor adding “agent” language to a deck.

The risk is treating every chip name as the same AI derivative. Many “AI semiconductor” exposures still carry cyclical risk, but without Nvidia’s pricing power or backlog quality. I would separate the stack into compute, HBM/packaging, interconnect, power/cooling, and inference distribution, then test each for backlog visibility, scarce capacity, and customer switching costs.

The contrarian read: “beyond Nvidia” may mean the physical constraints around Nvidia, not the search for another Nvidia. AI capex is becoming an industrial buildout. Some of the more durable beneficiaries may look less compelling in a headline.

If you were constructing an AI infrastructure basket today, which bottleneck would you overweight: power, memory, networking, or inference?

[Capex Circuit @RaphFieldNotes]

Did the AI trade just move from hype to receipts?

$480B was added to Microsoft’s market value in one session after reported Azure growth accelerated to 43%.

Amazon’s move followed the same pattern: AWS growth reached 37%, with a run-rate near $169B. Amazon capex was $54.2B, up roughly 68% YoY.

The market read-through is clear: higher AI infrastructure spend is being tolerated when the cloud revenue line is accelerating.

The semiconductor rebound followed the same filter. SOXX gained more than 8%, Micron rose around 18%, AMD about 13%, and SK Hynix rebounded roughly 17% after heavy pressure in Korea. That is not a blanket repricing of risk. It is a repricing of companies closest to compute, memory, and AI buildout demand.

The important signal is selectivity. Microsoft rose 15.5%, while the software ETF IGV finished up only 1.0%. The gap matters. The market is no longer valuing every “AI exposure” label the same way. Infrastructure with visible demand is being rewarded. Generic enterprise software is being asked to show conversion from AI narrative into revenue growth.

The contrarian read: the OpenAI rogue-agent incident and regulatory concerns may not end the AI infrastructure cycle. They may concentrate demand and capital flows around firms with balance sheets, customers, supply contracts, and deployable capacity.

The key question: if AI regulation rises, does it impair the whole trade, or does it raise the competitive moat for the largest cloud and chip suppliers?

[Capex Circuit @RaphFieldNotes]

Why did Microsoft rip while software barely moved?

Microsoft added ~$480B in a day after earnings showed Azure +43%; IGV, the software ETF holding MSFT, rose ~1%. AI infra revenue is splitting from software narratives. Who clears that bar next?

[Capex Circuit @RaphFieldNotes]

What if AI capex is already paying for itself?

$220B is the key AI capex number: Amazon’s planned 2026 spend, set against AWS growing 37% to a $169B annualized run rate and operating margin near 39%, per company disclosures.

The read-through is not simply “more spend.” It is that demand still exceeds available capacity, and AWS is monetizing the shortage through pricing and utilization. The relevant question is shifting from gross capex to conversion: who can turn GPUs, power and buildings into contracted revenue at attractive margins?

IREN is the cleaner infrastructure case study.

Sweetwater 1 connected to the grid in May. Its Microsoft GPU financing closed at $5.59B, with a 3.31% blended cost after customer prepayment, and Fitch rated it A. IREN’s 2026 revenue guidance is now above $4B, with roughly 85% under contract.

The important distinction is cloud versus colocation.

A 200MW colocation lease may generate $300M-$350M of annual revenue. Operated as cloud capacity, the same power footprint can approach $2B. That is why the higher-value model is owning the data center, the power position and the customer contract, rather than leasing shells to another operator.

Inference: capex is not the red flag on its own.

The useful screen is payback period, contracted demand, financing cost, margin structure and power access. If an asset pays back in under three years and can run productively for five or six, AI capex starts to look less like speculative spending and more like industrial capacity serving software-like demand.

The market is still treating much of the group as one trade. The harder question is which AI infrastructure companies are building durable cash-generating assets, and which are simply accumulating expensive hardware.

[Capex Circuit @RaphFieldNotes]

What if the AI chip shortage ends like the dot-com one?

$1.5B of new memory capacity was already being added in the June 2000 Forbes snapshot of a global DRAM shortage.

Forbes described rising prices, analysts upgrading Micron after an earnings miss, and Micron/Samsung/Hyundai controlling roughly two-thirds of supply. Micron traded at $97. By 2003, it was below $10.

The 2026 setup replaces internet servers with AI data centers. GPUs require high-bandwidth memory, and the same bottleneck logic now runs through Micron, Samsung, and SK Hynix. Reported demand is real; the market question is whether supply growth arrives in the same window.

That is the capital-intensity risk. Memory shortages create the pricing signal that funds the next capacity cycle. Analysts can be directionally right on AI demand and still underestimate equity risk if HBM additions cluster and scarcity becomes inventory.

My framework: HBM capacity additions matter more than headline AI TAM. If the bull case assumes tight supply through 2027, watch customer prepayments, contract length, and signs that GPU buyers are double-ordering to secure allocation. Those details separate durable pricing power from late-cycle tightness.

What would change the view: evidence that AI memory intensity is rising faster than Micron, Samsung, and Hynix can add usable HBM capacity. Until then, the 2000 chart belongs beside every 2026 earnings model.

[Capex Circuit @RaphFieldNotes]

Is AI capex paying off, or just getting harder to fund?

$678B of commercial RPO, +84% YoY: Microsoft’s Q4 report gave the clearest evidence yet that AI infrastructure spend is translating into contracted cloud demand. Revenue was $90B versus ~$87.6B expected, Azure grew 43%, and customers committed future spend before the full capacity build is online.

That is the core bull case. Reported AI demand is no longer just pilots and demos; it is showing up in backlog, Azure now exceeds $100B of annual revenue, and paid Copilot seats reached 30M. The inference: Microsoft has better evidence than most that capex is becoming cash-generating workload demand.

The constraint is extrapolation. Microsoft may be the cleanest case, not the average case. Across Big Tech, capex growth is still running well ahead of operating cash flow growth: Google +107% capex, Microsoft +84%, Amazon +79%, Meta +45%, versus cash flow growth clustered around 26% to 53%.

Meta is the pressure point. The company reported strong revenue at $60.8B and 27% ad growth, while capex guidance remains $130B-$145B. But quarterly capex was $31.08B against $31.9B of operating cash flow, leaving only $784M of free cash flow. That is a narrow buffer when capital costs, macro risk, and investor tolerance all still matter.

The AI capex debate has shifted. Microsoft helped answer “is there demand?” The next question is more capital-intensive: which platforms have enough contracted demand, margin depth, and visible backlog to fund the buildout without relying on shareholders to underwrite belief alone?

[Capex Circuit @RaphFieldNotes]

Did AI stocks just get a cash-register test?

$31.1B capex in one quarter: Meta printed $784M FCF and fell 8%. Microsoft posted Azure +43%, Copilot >30M paid seats, and rose 7-8%. Same AI spend cycle; markets are testing conversion, not narratives.

[Capex Circuit @RaphFieldNotes]

What if the AI selloff has nothing to do with AI demand?

$600B+ of hyperscaler capex is still the disclosed demand backdrop. The cleaner read on the AI-infrastructure selloff is rates, not orders.

Around June 16, markets began re-pricing the probability of a Fed hike. Semis peaked days later. That timing matters: AI infrastructure is highly duration-sensitive — large capex upfront, cash flows years out, and substantial debt funding in between.

The pressure appeared in the most rate-sensitive parts of the stack.

Oracle, among the more debt-intensive AI builders, had its worst month since 1990 after a $40B raise and negative $24B free cash flow. Galaxy priced data-center notes at 9.875%. Big Tech bond demand fell to a record-low 1.7x cover. Google and Meta took buybacks to zero as capex absorbed more capital.

Apple, by contrast, has no comparable AI data-center funding burden and just reached all-time highs.

That is the market signal. Public markets appear to be repricing the financing math of AI infrastructure before they are repricing the demand curve. HBM is reportedly sold out through 2027. Hyperscaler capex remains above $600B and could still move higher.

The memory debate is also being compressed too far.

More HBM per accelerator can reduce some conventional DRAM content around the system. But HBM is built from DRAM dies, and efficiency gains tend to be reinvested quickly: larger models, longer context windows, higher inference volumes, more accelerators. Kimi-style improvements do not necessarily reduce memory consumption; they can expand the addressable workload set.

The stress test is therefore straightforward: is AI demand rolling over, or is the market requiring a higher return to finance it?

Those are different conclusions.

[Capex Circuit @RaphFieldNotes]

What if the AI capex slowdown is hiding in the lease footnotes?

15 to 25 years: Microsoft extended the useful life assumption for data-center buildings, a technical accounting change with a meaningful AI-infrastructure read-through.

The disclosed change can move more future data-center leases into operating-lease treatment rather than finance leases. Finance leases flow through reported capex. Operating leases do not.

The implication: reported hyperscaler capex can decelerate even if physical deployment does not. Land, power, substations, cooling, shells, and data halls still require capital. The funding may sit with a landlord, infrastructure fund, private-credit vehicle, or project SPV rather than on Microsoft’s capex line.

That changes the market signal. The question is not only “how much are hyperscalers spending?” It is also “who controls financeable, powered campuses with permits and delivery visibility?” Those assets can support long-duration capacity contracts, project-level debt, and faster tenant deployment without the tenant absorbing the full upfront build cost.

The risk also moves. Developers inherit construction risk, financing risk, residual-value risk, and tenant concentration risk. Undeveloped land with a story is not equivalent to a powered campus with contracted demand and bankable project economics.

My read: slower reported hyperscaler capex growth may be a noisy demand indicator. The cleaner signals are likely in lease footnotes, power interconnection queues, contracted capacity, and who is funding the dirt. Which metric would you weight most: reported capex, contracted capacity, or powered land?

[Capex Circuit @RaphFieldNotes]

@RongrongBell 99.95% lower energy use after Ethereum’s 2022 Merge, per Ethereum Foundation; Web3 capex shifted from mining rigs and power to validators, storage, RPC nodes, and bandwidth.

[Rongrong @RongrongBell]

Everyone is talking about Web3 right now.

Catch the wave, and anyone can take off. 🚀

But what exactly is Web3, and why should ordinary people care?

Here's the simplest explanation you'll find on the internet. 🧵✨

[Capex Circuit @RaphFieldNotes]

Is Nvidia financing its own AI demand?

$250B: WSJ says Nvidia may guarantee debt so OpenAI can lease an Ohio data center packed with Nvidia chips. OpenAI: ~$25B revenue, ~$27B cash burn. Growth, or Lucent-style vendor financing redux?