The editorial argues that while equity markets value Anthropic at $150B+ based on revenue growth, bond desks correctly see a cash-burning company with single-supplier compute dependency (Nvidia), rented infrastructure (AWS/Google), concentrated enterprise customers, and well-funded competition. Ba1/BB+ is the appropriate rating for that risk profile, and the ~200bps spread penalty versus hyperscaler debt reflects a real structural disadvantage in funding the same compute.
The editorial highlights that Moody's and S&P explicitly did not treat Amazon and Google's strategic investments as quasi-equity support, nor did they credit Claude API revenue as durable. This represents a sharp divergence between how private equity markets and credit markets evaluate the same company — bond desks apply traditional metrics (free cash flow, asset coverage, revenue durability) rather than growth narratives.
By submitting the story with the headline 'Anthropic assigned highest junk credit rating in June,' the submitter elevated a detail that FT had buried inside a broader piece on AI capex financing. The 234-point score suggests the HN community agreed this was a signal worth isolating from the broader AI financing narrative.
In June, Moody's and S&P quietly assigned Anthropic its first corporate credit ratings ahead of a debt raise: Ba1 from Moody's and BB+ from S&P — the highest rung of speculative grade, one notch below investment grade. The Financial Times reported the detail buried in a broader piece on how private credit and public bond desks are pricing the AI capex wave. Anthropic is now the highest-rated junk credit in AI, which is another way of saying it is still junk.
The context matters. Anthropic reportedly crossed a ~$5B annualized revenue run-rate earlier in the year and is raising at a valuation north of $150B. By the numbers that drive equity rounds, that is an investment-grade story. By the numbers that drive bond desks — free cash flow, asset coverage, revenue durability, customer concentration — it is not. The rating agencies are looking at a company that burns cash to buy compute from a single supplier (Nvidia), rents that compute through a second party (AWS, with Google in the mix), serves a customer base dominated by a handful of enterprise contracts, and competes with a well-funded incumbent (OpenAI) and the hyperscalers themselves. Ba1/BB+ is what that risk profile earns in a market that just repriced duration.
The spread tells the rest of the story. Investment-grade AI-adjacent paper — Microsoft, Oracle's recent OCI-backed issuance, Meta's data-center bonds — trades inside 150bps over Treasuries. Ba1 paper for a cash-burning growth company lands north of 300bps, often closer to 400bps once you account for call protection and covenant quality. On the same compute, a hyperscaler's balance sheet funds it 200+ basis points cheaper than Anthropic's does.
The interesting part isn't that Anthropic got rated junk. The interesting part is what the rating agencies refused to credit. They did not treat the Amazon or Google strategic investments as quasi-equity support. They did not treat the Claude API revenue as sticky in the way they treat Azure or AWS contracts. They did not treat model IP as a durable asset you can lend against — which is the right call, because model weights depreciate on a six-to-twelve-month cycle and have essentially zero liquidation value. In bond-desk language: AI labs are capex-heavy businesses with software-company balance sheets and consumer-electronics-grade product cycles.
This matters because of what it implies about the structure of the AI build-out. The hyperscalers have been fronting an enormous amount of capex — Microsoft, Google, Meta, and Amazon are collectively on track for ~$350B in 2025 — and funding it through investment-grade debt and operating cash flow from products that mostly aren't AI. The labs themselves cannot do that. They have to raise equity at ever-higher valuations or issue debt at junk spreads. Anthropic's rating is the first time we've seen a lab's cost of debt capital priced in the open, and the number is: expensive.
Compare the structural options. OpenAI has effectively outsourced its balance sheet to Microsoft and SoftBank. xAI is funding compute through private credit at even wider spreads — reportedly double digits on some tranches. Mistral and Cohere are too small to tap the bond market at all. Anthropic going to the public bond market at Ba1 is actually the healthy version of this story — the alternative is private credit at usury rates or more dilutive equity.
There's also a reflexivity problem the ratings surface. If Claude usage growth slows for even a quarter, the Ba1 rating is on review for downgrade. A downgrade to Ba2 or Ba3 would push the next raise another 100-200bps wider, which cuts into gross margin on API calls, which pressures the growth number that justifies the valuation that supports the next equity raise. The agencies are explicit about this feedback loop in their rationale. It is why the rating is Ba1 and not Baa3: the equity story and the credit story disagree, and when they disagree, bond desks win because they are the ones setting the price of the next dollar in.
Community reaction on HN was mostly focused on the valuation-vs-rating gap — "how can something worth $150B be junk?" — but that framing misses the point. Equity values optionality. Debt values the floor. A company can be worth $150B in equity because of a 10% chance it's worth $1.5T, and simultaneously be rated junk because of a 20% chance it's worth zero in five years. Both are true at once, and the Anthropic rating is the first clean data point we have on where the market puts those probabilities.
If you're building on top of a frontier lab's API, this is the number that will quietly set your 2026 pricing. API prices are not set by inference cost — they are set by inference cost plus the amortized cost of the next training run, discounted at the lab's cost of capital. Anthropic's cost of capital just got a public number attached to it, and it's higher than most engineering leaders are modeling. Expect the per-token economics on Claude to compress less aggressively than the hardware curve would suggest, because the lab needs the margin to service debt and fund Claude 5.
For platform teams, the practical read is: diversify your model provider dependency the same way you'd diversify a single-supplier database contract. Not because Anthropic is going anywhere — Ba1 is a healthy credit, not a distressed one — but because single-provider lock-in to a Ba1 counterparty is a different risk posture than lock-in to AWS. Price your fallback. Keep your prompts portable. If you're on an enterprise contract, read the MAC and termination clauses, because the bond covenants on this kind of paper usually force the issuer to protect enterprise revenue above all else, which is good for you but worth knowing.
For anyone running an AI startup burning venture money on API calls: the lab's cost of capital is now a floor under yours. If Anthropic is paying ~400bps over Treasuries for compute, you are not getting a sustainable discount on the output of that compute. Business models that only work at 2024 API prices are not going to be rescued by hardware deflation alone.
The next data point to watch is what happens when OpenAI finally issues standalone debt — not Microsoft-backed, not SoftBank-wrapped, but on its own balance sheet. The over/under in bond-desk chatter is Ba2, one notch wider than Anthropic, because of the governance overhang and more aggressive capex. If that print comes in at Ba1 or tighter, the market is saying AI labs are converging on a stable credit profile and the build-out gets cheaper from here. If it prints at Ba3 or wider, the market is saying the capex arms race has gotten ahead of the revenue, and 2026 API prices have further to fall before they stabilize. Either way, the ratings page just became required reading for anyone building on frontier models.
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.