The coupon expired: AI's capacity bill is arriving as a memory price, and duration decides who pays

A four-year-old CPU socket is the clearest priced signal available on what AI capacity costs the rest of the stack. TechSpot's reading is blunt: the RAM crisis has made the newest hardware unaffordable enough that components from four years ago are storming the sales charts ^[1]. Intel's Robert Hallock, vice president and general manager of its enthusiast channel business, told Tom's Hardware that the company will keep selling Raptor Lake and means to stabilise supply of its 10-nanometer LGA 1700 parts "for years to come" ^[2], and TechSpot notes reports of a third Raptor Lake refresh in the first half of 2027 alongside at least two motherboard vendors raising output of aging LGA 1700 boards ^[3]. The cause is memory. DDR5 has climbed steeply over the past year on shortages TechSpot attributes to AI data centres, and what makes a 2022 chip worth buying in 2026 is that it will run DDR4 ^[4].

What the ability to re-issue is worth

AMD has put the 2022 Ryzen 7 5800X3D back on sale, and the parts that would otherwise beat it, the 7500F and 9800X3D, all require DDR5 ^[5]. Nvidia has revived the 12GB RTX 3060 as a temporary substitute while GDDR7 tightness raises RTX 5000-series prices and delays the Super cards ^[6]. That is the same manoeuvre three times: when an input reprices, whoever owns a finished design and a live socket can reach back into a product they have already amortised. The buyer's equivalent move is much thinner. TechSpot describes it as paying inflated DDR4 prices or holding the RAM already installed ^[4].

GitHub shows what the move looks like when you can simply buy the increase. Its CTO Vlad Fedorov's account of the August 17 outage, a failure of 7 hours 47 minutes that took down github.com, authentication, Actions, the APIs, pull requests, issues and Copilot after a critical component in the Central US data centre failed to scale under record traffic ^[7], comes with capacity figures: more than 3 million CPU cores and 120 petabytes of high-speed storage added ^[9]. Azure went from 12% of platform load in May to 58%, carrying half of all Git operations ^[10], while monthly commits went from 1.4 billion in April to 2.9 billion in August ^[11], slightly better than a doubling in four months ^[1]. Multiply share by volume and the Azure-hosted portion goes from roughly 168 million commits a month to roughly 1.68 billion, about ten times as much, though the share is measured in May against April volume, so read it as an order of magnitude rather than a measurement ^[2]. GitHub absorbed a capacity squeeze by adding cores and moving load to a hyperscaler it can call internally. That option is not on the menu for a team whose infrastructure arrives as an invoice.

The bill that arrives as a failure rate

Running denser has a metallurgical price, and it is now quantified. Christoph Siemroth of Essex and Yeomyung Park of Sungkyunkwan re-crunched twelve years of Backblaze data, 443,156 drives and more than 1.66 million drive-years from 2013 to Q2 2025, in a peer-reviewed IEEE study; matched on age, capacity, form factor and temperature, HGST fails at about 41% of Seagate's rate, WD at about 52%, and Toshiba at 107% ^[13]. The number that matters for anyone packing more into the same room is thermal: each additional degree Celsius of average drive temperature raised failure rates by 2.1%, compounding to roughly 23% across ten degrees ^[15]. Backblaze put its own fleet-wide annualised failure rate for 2025 at 1.36% ^[16]. Apply the study's coefficient and ten degrees of surrendered thermal headroom takes that to about 1.67%, an extra 3.1 dead drives per thousand per year ^[3]. That is the cooling bill restated as replacement hardware, payable by whoever owns the metal.

The study also complicates the hold-what-you-have strategy that desktop buyers have adopted. Each extra terabyte of capacity cut failure rates by about 3.4%, which the authors attribute partly to helium enclosures and newer head fabrication ^[15], so the drives least likely to fail are the ones you would have to buy. Ageing is not brand-neutral either: Toshiba's monthly failure rate more than quadruples past 60 months, running at five to six per thousand drives per month between 50 and 100 months while the other three brands generally stay at or below 2.5 ^[17]. And the estimates are honest about their reach, with only 0.52% of sampled drives running longer than a decade ^[18]. The ranking's winner is the one nobody can order: WD acquired HGST in 2012 and wound the brand down ^[14].

The layer with no inventory

Now the top of the stack. Gergely Orosz spoke with almost 20 engineering leaders on or considering a career break and reports that 6 out of 10 CTO-level people he asked said they are on the way out ^[19]. The stated pressures are budgetary and specific: engineering cost cuts of 20% to 50% including job cuts, and pressure on business results as AI coding bills rack up ^[20]. Note what is missing from that description. Intel can name a socket and a year ^[2]. The study can name a coefficient per degree ^[15]. The application layer's AI cost appears in these accounts only as a bill that grows, with no unit price attached to it ^[20], which is the visibility gap in one line.

Duration pricing shows up even at the smallest scale. Docker Desktop is free personally and professional plans start at $11 a month, or $9 a month billed annually ^[23]: $132 against $108, so committing a year buys an 18% discount on the monthly rate ^[5]. Meanwhile about 36% of Docker development is still done on local machines, per the 2025 Docker State of Application Development ^[24], which means that share of the compute is running on hardware someone already bought, and replacing it is precisely what has become expensive ^[1].

Self-help at the code layer does not close the gap. Go 1.27's size-specialised allocation cuts small-object allocation costs by up to 30% and improves overall performance by about 1% for allocation-heavy programs ^[34]. That is a real win, and it is measured on a different base than a 20% to 50% cost reduction target ^[20]. No runtime release arithmetically meets a mandate of that size, which is why the mandate lands on people.

The equity side is worse for being imprecise. DevOps'ish, summarising the Orosz piece, cites one CTO holding 2% at a company that raised $110M on a 2x preference, "meaning a $210M exit before the stock is worth anything", and notes the original is paywalled partway through ^[22]. Two times $110M is $220M, ten million above the figure as printed ^[4]. It is a small discrepancy in a secondhand summary, and it sits on top of the single number that decides whether a decade of work pays.

Retries are a cost transfer

The clearest example of the application layer moving its own costs upstream is in GitHub's own post-incident account: Copilot recovered last because a client-side retry loop pushed more traffic into a system that was trying to come back ^[8]. Nobody billed the clients for that decision; the recovery paid for it.

Those settings are usually inherited rather than chosen. The Kubernetes defaults get spelled out plainly in the ngrok write-up: failureThreshold of 3, successThreshold of 1, a 30-second termination grace period, and CrashLoopBackOff starting at 10 seconds and doubling to a five-minute ceiling ^[29]. The author also found and reproduced, on k3s, minikube and kind, a bug introduced in v1.35.0 where liveness probes fire before the startup probe has succeeded, the exact failure mode that makes a slow-starting service restart forever ^[30]. Short visibility is not a metaphor here. It is a specific version number in which the thing that decides whether your container gets killed was wrong.

Buying back control costs four engineer-years

The parties in this material that dislike depending on someone else's substrate are all doing the same thing, and the price is visible. Encore has run every build inside a Firecracker microVM since mid-2022, but Firecracker drives KVM and needs /dev/kvm, which no Mac has, and its maintainers turned down a working proof of concept built on Apple's Virtualization.framework and said they do not plan to support macOS any time soon ^[26]. So for four years, working on the build system meant working on it somewhere else, until Encore built crackling, one microVM API driving Firecracker on Linux and Apple's hypervisor on macOS, which required rebuilding much of the Linux image toolchain to run on macOS ^[27]. Before that, every engineer had a user on one shared build box, a port kept in a gitignored CUE file so nobody collided, and a homemade layer-to-squashfs conversion piped over SSH because no existing tool turned Docker layers into a block device Firecracker could boot ^[28].

Docker has made the same call one layer down, replacing libkrun on Apple, Hyper-V and WSL on Windows, and KVM and QEMU on Linux with its own Docker VMM, in beta with Desktop v4.86 on Windows and macOS and Linux at general availability, on the stated grounds that owning the full stack lets it tune every part of the engine for container workloads ^[25]. OpenDepot makes the argument in a README: a self-hosted OpenTofu and Terraform registry implementing both registry protocols, for complete control over distribution, versioning and storage without relying on the public registry ^[32].

Insourcing your substrate, then, is a hedge available to whoever can spend engineer-years, and Encore's account prices it at four years of friction plus a rebuilt image toolchain ^[27]. Everyone else keeps buying capacity they do not control, on the shortest notice in the stack. The vendors holding sockets and designs will spend 2027 selling older parts under newer names ^[3]; the layer billed monthly gets to pick which inherited defaults to keep ^[29], and to find the 20% to 50% somewhere that is not silicon ^[20].