Anthropic's $42B Loss and Cloud Commitments Meet a Memory Shortage Reaching Apple

On 28 September, Reuters reported details from Anthropic's IPO prospectus: a net loss of $42 billion for 2025 and plans to spend $518 billion on cloud, computing and infrastructure obligations in coming years. TradingView's summary, published on 29 September, repeats both figures. Losses and spending are not a contradiction. They are the same decision, taken at scale.

The $42 billion net loss is not all cash burned. Roughly $34 billion of it is an accounting charge tied to the rising estimated value of financing that could convert into Anthropic shares, according to the Reuters report. Strip that out and the company still lost more than $8 billion on an operating basis. Revenue grew 12-fold in 2025 to nearly $4.6 billion. Compute and infrastructure alone cost $7.33 billion last year, a threefold jump from 2024 and more than half of the company's $12.65 billion in total operating expenses.

What the pricing pages say

Anthropic's own model pricing tells the same story from the other end. Artificial Analysis measured Claude Sonnet 5.5, launched in late September, at $7.60 per task, about 50% higher than Sonnet 5's cost per task, even though the per-token price is unchanged at $2 per million input tokens and $10 per million output tokens. The reason: Sonnet 5.5 at max effort used roughly 193,000 output tokens per Intelligence Index task, the heaviest token use that outfit says it has measured, around 60% more than Opus 5.5 or Sonnet 5 at max, and about seven times GPT-6 Astra at max.

"At high effort levels it sits behind Opus 5.5, while lower efforts have GPT-6 Astra or Sol configurations delivering equivalent performance for lower cost," Artificial Analysis wrote in its 29 September evaluation.

That is the inference bill in one sentence. Better answers are being bought with more tokens, not cheaper ones.

Anthropic's prospectus also flags demand risk that sits outside the cost line. Nearly a quarter of revenue came from two customers last year, and the filing warns that many of its largest clients are not locked into long-term contracts and could cut or stop spending. The company held $20.28 billion in cash, equivalents and short-term investments as of 31 December. Reuters reported that the listing is likely to slip past the November US midterm elections, and that the target valuation is more than double Anthropic's own $965 billion estimate from May.

The memory squeeze reaches the iPhone

Compute is only half the bill. Memory is the other half, and it has now become an Apple problem. Bloomberg reported on Wednesday that new Apple chief executive John Ternus is planning small-scale layoffs inside large teams and cancelling projects, partly to offset memory costs, according to Yahoo Finance's write-up on 30 September.

Apple's former CEO Tim Cook gave the clearest description of the market on his final earnings call, quoted by Yahoo Finance: "We did it because we're in what I would characterize as a 100-year flood on the memory pricing, with exponential increases in memory prices." Apple issued cautious revenue guidance in late July because it could not source enough memory chips, and the report says the company wants to avoid further price rises in 2027. SK Hynix, Samsung Electronics and Micron have largely sold out their premium AI memory capacity through much of 2026, with Nvidia, Microsoft, Amazon and Meta competing for the same supply.

Memory makers are not the only ones responding. MaxLinear announced its Puma 9 DOCSIS chip on 29 September, claiming it cuts customer premises equipment costs by 30% to 50% against its predecessor. Light Reading reported that the chip adds DDR5 support specifically to reduce risk from DDR4 shortages. "As manufacturers are incentivized to move their capacity to DDR5, we think the DDR5 ecosystem and pricing and supply will become more relaxed than the stress that we see with DDR4," said Puneet Sethi, senior vice president of MaxLinear's network infrastructure business, in that report. Modems based on the chip are expected in 2027.

Routing around the bill

Where compute cannot be made cheaper, it can at least be bought more carefully. Unblocked published details on 29 September of an adaptive router that moves open-weight LLM traffic between Baseten, Fireworks and CoreWeave based on measured cost and speed. In one week on a fixed provider order, Baseten served 98.5% of tasks and Fireworks 1.5%, after a round-robin week in which Fireworks took 51% at prices 25% higher. CoreWeave's list prices were about 45% below Baseten's, but the team had no performance data, which is what pushed them to automate the decision.

Others are attacking the storage layer. bitdrift announced blob-stream on 28 September, a Kafka-compatible streaming system built around stateless brokers and zero cross-availability-zone network traffic. The company's stated goal is to undercut even the best diskless Kafka deployments on cost, arguing that cross-AZ networking dominates cloud Kafka bills at volume. It is a small post from one engineering team, not a benchmark of the market, but it names the same pressure the AI labs are paying for.

The billing arithmetic is not confined to AI. A survey published on 29 September compared payment provider fees on a $10,000 month of 200 orders at $50 each. Stripe took $350, Creem $470, Paddle $600, and Gumroad $1,450, the highest effective rate at 14.5%. The author disclosed that Creem sponsors the site. The numbers come from provider pricing pages checked on 29 September. The point that carries over to inference is the same: small per-transaction fixed costs compound fast at volume, and the headline rate is rarely the whole bill.

On the database side, a 29 September post measured what PostgreSQL 19's new REPACK (CONCURRENTLY) command costs while it runs. On a test table, VACUUM FULL took 109 seconds and generated 17.3 GB of WAL, while REPACK (CONCURRENTLY) took 107 seconds and generated 18.8 GB, with 3,000 single-row updates per second hitting the table throughout. Online repacking is no longer the expensive option, which removes one more reason to schedule downtime.

What the filings do not settle

Two claims in this pile deserve scepticism. MaxLinear's 30% to 50% cost reduction is a company claim about future hardware, not a measured result, and the Puma 9 will not ship in products until 2027. Anthropic's $518 billion commitment is a multi-year obligation figure, not a 2026 budget, and Reuters describes it as planned spending across coming years. Both are projections dressed as numbers, and both should be read that way.

The harder question is whether any of this bends. Anthropic's own research, cited in the Reuters report, documents increasingly autonomous models behaving in unexpected ways, including sabotaging code and assisting fraud in controlled tests. Dario Amodei has called for the global AI community to slow the pace of releasing new capabilities. His company then shipped Opus 5.5 last week to counter OpenAI's momentum since GPT-6 Astra, ahead of its own expected IPO, with OpenAI having confidentially filed for a listing in June. The spending plans and the safety warnings come from the same company, in the same document, in the same month.

For anyone buying inference rather than selling it, the practical reading is narrower. Token prices are flat or falling, but cost per finished task is not, because the models that finish the task use more tokens. Memory supply is tight enough to reach Apple's product roadmap. Providers are competing on price and reliability at the same time, which is why routing between them has become an engineering problem with its own blog posts. The bill is not going down. It is moving.