Anthropic's Compute Bill Meets a Memory Shortage and Rising GPU Prices

On 28 September Reuters reported that Anthropic's IPO prospectus shows a net loss of $42 billion in 2025 and planned spending of $518 billion on cloud, computing and infrastructure obligations in coming years. The company spent $7.33 billion on compute and infrastructure last year, a threefold rise from 2024, according to the same document. TradingView's write-up of the Reuters report adds that Anthropic's $12.65 billion in total operating expenses means compute alone accounted for more than half of them. The company ended the year with cash, cash equivalents and short-term investments of $20.28 billion as of December 31.

The bill is not just GPUs

Memory is where the pressure is now visible.

Yahoo Finance reported on Wednesday that Apple's incoming CEO John Ternus is planning small-scale layoffs and canceling projects, citing a Bloomberg report, and it ties the move to a memory shortage raising component costs. The piece quotes former CEO Tim Cook on his final earnings call: "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." That shortage has a supply side. Yahoo Finance notes that SK Hynix, Samsung Electronics and Micron have largely sold out their premium AI memory capacity through much of 2026, as Nvidia, Microsoft, Amazon and Meta race to build AI infrastructure. It adds that suppliers expect memory supply to remain constrained into 2027.

"We're pioneering manganese-rich battery technology to unlock premium range and performance at an affordable cost, especially in electric trucks."

Not every hardware cost is climbing.

MaxLinear said on 29 September that its new Puma 9 DOCSIS chip cuts customer premises equipment costs by 30% to 50% versus its predecessor, according to Light Reading. The chip adds DDR5 memory support to ease the DDR4 squeeze. Puneet Sethi, SVP of MaxLinear's network infrastructure unit, told Light Reading that as manufacturers move capacity to DDR5, "we think the DDR5 ecosystem and pricing and supply will become more relaxed than the stress that we see with DDR4." The same memory logic is moving through autos. Electrek reported on 29 September that Ultium Cells, GM's joint venture with LG Energy Solution, will mass-produce prismatic LMR battery cells at its Spring Hill, Tennessee plant, claiming 33% higher energy density than LFP at comparable cost. GM's Kurt Kelty said the chemistry lets the company "leapfrog today's more affordable chemistries." Upgrades start later this year and are expected to finish in 2028, adding 500 jobs.

What inference actually costs per task

Artificial Analysis published a review of Anthropic's Claude Sonnet 5.5 on 29 September, measured on a pre-release build. It scored 56 on the Artificial Analysis Intelligence Index, two points behind Opus 5.5 (max), while using roughly 193k output tokens per index task, about 60% higher than Opus 5.5 or Sonnet 5 and roughly 7x GPT-6 Astra. At Anthropic's $2/$10 per million input/output token pricing, that works out to $7.60 per task, about 50% higher than Sonnet 5's cost per task.

That is the tension operators are pricing in. Unblocked described on 29 September an adaptive router that moves traffic across Baseten, Fireworks and CoreWeave, all serving GLM 5.2. It found CoreWeave list prices about 45% lower than Baseten's, but says it had no production data on CoreWeave's reliability, so it initially could not place the provider in a fixed order. The router measures cost and speed per task, not per model call.

Underneath the routing layer, streaming infrastructure is being re-costed too. bitdrift announced blob-stream on 28 September, a Kafka alternative whose stated first goal is zero cross-availability-zone network traffic, which it says dominates cloud Kafka costs at volume. Its brokers are stateless with no local storage.

Database maintenance is another line item. boringSQL measured PostgreSQL 19beta4's new REPACK (CONCURRENTLY) on 29 September. On its test, VACUUM FULL took 109 seconds and generated 17.3 GB of WAL, while REPACK (CONCURRENTLY) took 107 seconds, generated 18.8 GB and used 18.7 GB of peak extra disk. pg_repack took 162 seconds and generated 33.5 GB of WAL; pg_squeeze took 165 seconds and 18.8 GB.

Who pays for verification

Agent work creates a new cost category. Append Only published on 29 September a 55-day pilot of its record-keeping system, putting the cost at roughly $15 a day at list price, $829.12 total, with 557 verifications and 43 incidents. It says 98% of context was served from cache and a fresh instance needed $1.04 to orient from the record alone. These are the vendor's own numbers from its own pilot, not an independent audit.

Twin markets a different model: unlimited agents on monthly subscription pricing, with 257 templates advertised on its site on 29 September. Both approaches are attempts to make per-task costs predictable, a problem that gets harder as token use per task rises.

Payment overhead is a reminder that AI margins sit inside ordinary business costs. Flaviocopes calculated on 29 September that on $10,000 a month in sales across 200 orders at $50, Stripe takes $350, Creem $470, Paddle $600, Polar's free plan and Lemon Squeezy $600 each, and Gumroad $1,450. The author discloses that Creem sponsors the site. Card processing and merchant-of-record status explain most of the spread.

There is a wider argument about how prices get set at all. The American Prospect reviewed Lindsay Owens's book Gouged on 29 September, noting her earlier work with Groundwork Collaborative, Consumer Reports and More Perfect Union found roughly 75% of items in identical Instacart baskets varied in price between shoppers at the same time. The review says the critique spurred an FTC investigation and that Instacart later disavowed the technology.

Two other pieces in the dossier point at how the compute market is being repriced. Captain ACAB's public dataset of law enforcement settlements lists a $36,000,000 wrongful death settlement approved in Alameda County on 2026-05-12 and an $8,150,000 wrongful death settlement in Orange County on 2026-05-19, a reminder that public costs often sit far from the technology that produces them. Pedro Santa Clara's 29 September essay on asset pricing notes that the field's canonical history was assembled in the 1960s by people who needed a lineage, a caution against reading today's AI capex stories as settled arithmetic.

The near-term question is not whether AI compute is expensive. It is whether memory supply, routing and verification costs fall fast enough to matter before the next round of commitments lands.