CoreWeave Credit Fears and Slower Anthropic Growth: Is the AI Infrastructure Boom in Trouble?

AI infrastructure skepticism has a real financial basis, but it does not yet prove that demand for AI chips and memory is collapsing. CoreWeave’s funding burden, Anthropic’s reported growth deceleration, memory supply contracts and the changing economics of inference describe different parts of the same investment chain.

Evidence checked October 19, 2026, Korea time. The interpretation and scenarios are the author’s; company statements, public filings and reported private-company figures are distinguished below. This is industry analysis, not a recommendation to buy or sell securities.

Contents · the terms behind the headlines

Credit risk — whether a borrower can meet its obligations.
Revenue growth — recognized sales, not an annualized headline.
LTA / SCA — multi-year supply arrangements with specific terms.
Chip commitments — plans, agreements and delivered systems are different.
KV cache — stored attention keys and values used during inference.
Outlook — conditions that would strengthen or weaken the thesis.
Physical AI — robotaxi and humanoid opportunities beyond chatbots.

1. CoreWeave: strong demand does not eliminate credit risk

Credit risk is the risk that contractual payments cannot be met when due. An AI cloud can have valuable customers and still face a mismatch between spending on capacity today and collecting revenue later.

CoreWeave’s June-quarter 2026 results show the tension: revenue of $2.575 billion, net interest expense of $640 million and a GAAP net loss of $626 million. Its approximately $104 billion June-end revenue backlog includes amounts beyond remaining performance obligations and depends on delivery and service-availability requirements. CoreWeave Q2 results and June 30 Form 10-Q.

Those are not contradictory signals. Backlog indicates contracted future activity, not cash available for the next interest bill. Adjusted EBITDA is also not free cash flow after infrastructure investment. The useful credit questions are refinancing terms, customer concentration, deployment timing and whether usable capacity starts earning before financing costs accumulate.

The investment chain has several separate gates

1. Customer signs a commitment — contractual quality and counterparty funding matter.

2. Supplier builds chips, memory and racks — qualification and delivery still matter.

3. Data center energizes capacity — power, cooling and utilization still matter.

4. Provider collects cash — operating costs, interest and replacement investment still matter.

Original explanatory diagram, not a company cash-flow forecast. A long-term contract strengthens the first gate; it does not automatically clear all four.

2. Anthropic: slower growth is not falling revenue

An October 15 Euronews report, citing investor presentations reported by Bloomberg, put Anthropic’s July–September revenue at $16.8 billion, versus $11.5 billion in the preceding quarter: roughly 46% growth. The reported pace was slower than the previous quarter’s more-than-doubling. These are reported private-company figures, not a public audited quarterly filing independently inspected for this article.

A business can keep expanding quickly and still disappoint expectations embedded in financing plans or valuations. Annualized run-rate revenue extrapolates a recent pace; it is not revenue already earned over a year. Nor can two quarterly totals reveal a reliable month-by-month trend. The defensible concern is whether monetization and cash generation can support commitments—not that these reported numbers show sales collapsing.

3. Memory LTAs are substantial—but “everything sold through 2028” is too broad

An LTA is a long-term agreement. Its protection depends on the covered product, committed volume, price mechanism, deposits, enforcement and delivery conditions. HBM, conventional DRAM and NAND are not one interchangeable market.

Supplier evidence	What it does—and does not—establish
Micron · September 30, 2026	Its FY2026 Q4 remarks, page 3 describe 26 multi-year take-or-pay SCAs, estimated at over 35% of revenue through 2030; some extend into 2031. Customer financial commitments total $32 billion, mostly deposits. This is not 100% of all future output.
SK hynix · July 29, 2026	Q2 statement: LTAs finalized with around 10 customers. That confirms multi-year contracting, not a public company-wide, product-by-product “fully booked to December 2028” schedule.
Samsung · July 25, 2026	Broadcom collaboration announcement: an MOU covers memory and foundry cooperation, including HBM. An MOU and broad collaboration estimate are not a disclosed, binding schedule for all memory capacity.

My reading is constructive but selective: stronger supply assurance and customer deposits can reduce a memory maker’s planning risk. They cannot remove counterparty risk, technology qualification, manufacturing yield, or the possibility that new capacity changes pricing later. Public evidence reviewed here does not establish that every major memory maker has all output locked up through end-2028.

4. Chip agreements show a pipeline, not a single guaranteed revenue pool

NVIDIA’s September 2025 OpenAI announcement was explicitly a letter of intent for at least 10 GW of systems. OpenAI and Broadcom’s October 2025 10 GW collaboration described existing co-development/supply agreements and a rack-deployment term sheet, targeting second-half 2026 through end-2029. Anthropic’s April 2026 Google/Broadcom agreement expects next-generation TPU capacity to begin coming online in 2027.

These announcements support a multi-year buildout thesis. They are not equivalent legal instruments, and their gigawatts must not be added as independently verified installed capacity or converted into guaranteed HBM orders. GPU and custom-accelerator platforms can also compete for the same workload. Supplier revenue depends on actual products, acceptance, deployment and payment.

The reverse exposure matters too: NVIDIA’s FY2027 Q2 filing discusses infrastructure-related commitments and credit support, as well as inventory/purchase-obligation risks. A supplier that helps finance a customer may benefit from demand while also retaining some ecosystem risk. Counting the investment and the customer’s planned purchases as two independent proofs of end-user demand would be misleading.

5. LLM progress changes both demand and efficiency

There is no defensible universal “LLM improvement rate” or annual KV-cache growth rate. A useful recent counterexample to “better models always consume more per task” is Anthropic’s September 22 Opus 5.5 release: it claims improved capability and lower running cost. Those are vendor-reported results, not a benchmark reproduced here; savings on selected tasks are not an industry-wide reduction in total compute.

KV cache stores the attention keys and values needed to continue generating without recomputing the whole history. Longer active context and more simultaneous sessions increase conventional full-attention cache requirements; fewer KV heads, lower precision, bounded windows and reuse can reduce them. Maximum advertised context is not the average context actually used.

Illustrative full-attention cache model
KV bytes ≈ 2 × layers × KV heads × head dimension × retained tokens × bytes per value × concurrent sequences

Assume 32 layers, 8 KV heads, head dimension 128 and 2-byte values. This gives 128 KiB per retained token per sequence. It excludes weights, activations, allocation overhead and shared-prefix savings; it is not a frontier model’s disclosed architecture.

KV cache grows with the workload—not the calendar

1. 32,768 tokens × 1 sequence → 4 GiB of raw KV data.

2. 131,072 tokens × 1 sequence → 16 GiB: 4× the cache, or +300%.

3. 131,072 tokens × 4 sequences → 64 GiB: 16× the starting total.

Original hypothetical arithmetic, executed in Node.js—not a GPU or LLM benchmark. GiB = 2³⁰ bytes. These are total raw cache sizes before sharding; no claim that one GPU can host all of them.

The tensor layout is grounded in Hugging Face’s cache explanation. Its cache-strategy documentation also describes bounded cache growth, quantization and CPU offloading. Moving cache to CPU DRAM does not make it HBM, and SSD storage is not a drop-in substitute for low-latency GPU memory. Offloading changes transfer costs and latency, not just capacity.

The investment implication is a balance, not an automatic multiplier. Useful agent sessions, multimodal workloads and longer histories could raise total retained state. Compression and cheaper models push the other way. If context and concurrency each rise 4× while bytes per value halve, this toy model still grows 8×—but that is a conditional workload scenario, not a forecast of memory sales.

6. What would make the growth thesis stronger—or weaker?

Scenario · not a probability forecast	Evidence to watch
Durable expansion	Paid usage and retention improve; deployed capacity earns cash; deposits convert into accepted deliveries; financing remains serviceable. Chips, HBM, CPU DRAM and storage benefit differently.
Growth with financial shakeouts	AI usage grows, but expensive borrowers refinance poorly or delay projects. Efficient operators and selected suppliers can grow while some equity or debt investors lose money.
Demand disappoints / excess capacity	Customer budgets weaken, contract deliveries slip, utilization and realized rental prices fall, or supply additions outrun profitable usage. Longer contracts cushion some timing risk, not every loss.

My base interpretation is that a financing stress test is a better description than proven technological exhaustion. But that is an interpretation, not a measured probability. Watch cash conversion, deployment milestones, contract changes and realized pricing together. Rising token counts alone cannot prove that a rack earns an adequate return; a large backlog alone cannot prove that a borrower can refinance.

7. Robotaxis and humanoids could open the next large market—if commercialization works

Tesla’s June-quarter 2026 filing describes an expanding Robotaxi service, Cybercab production beginning in the first half, and preparations for large-scale Optimus production. These are different maturity stages. The filing does not establish that general-purpose humanoids already form a mature mass-market business.

The upside scenario is bigger than selling more chatbot subscriptions. Reliable autonomous transport could address paid mobility; capable humanoids could address repeatable physical work. Training, simulation, fleet learning, onboard inference, memory and sensing would then form a broader infrastructure chain. This is an architectural and commercial scenario—not a Tesla sales forecast, and not every action would become a cloud LLM request.

Safety, regulatory permission, intervention rates, manufacturing yield, maintenance, insurance and cost per useful ride or productive hour decide whether that opportunity becomes an enduring market. They also determine which chip and memory architectures win. Physical AI may open another enormous market, but its possible future cash flows cannot repay a cloud provider’s interest bill today.

Bottom line: take credit concerns seriously without treating them as proof that AI has stopped advancing. Contracts and inference workloads support a plausible long-term growth path; profitable deployment and cash collection determine who captures it. Robotaxi and humanoid commercialization would add a new growth engine—conditionally, not inevitably.

Continue this English reading thread

Revenue: why run-rate and recognized revenue differ — background on metric boundaries, not a current Anthropic data source.
Hardware: NVL72 HBM4 and SOCAMM2 demand scenarios — configuration arithmetic versus real orders.
Serving: prefix caching and the memory trade-off — why compute savings and memory demand can coexist.
Measurement: TTFT, TPOT, ITL and goodput — test the useful service delivered, not just peak token output.

Appendix · boundaries and source trail

HBM is high-bandwidth memory close to an accelerator; CPU DRAM supports a different memory tier; NAND flash is persistent storage. A gigawatt measures power, not delivered tokens, chip counts or supplier revenue. Take-or-pay strengthens a purchase commitment but is not immunity from default or delivery disputes.

Methods: desktop arithmetic only, Node.js v24.14.1 on Windows; no paid model call, hardware measurement or trading-price forecast. The Anthropic quarter figures are explicitly attributed reporting. No inference is made about individual monthly revenue, all suppliers’ undisclosed contracts, an industry KV-cache CAGR or a numerical robotics TAM.

이 블로그 검색

Deoks