Why iFixit's Steam Machine Guides Expose the Parts Gap

The bench needed a hair dryer to get inside a pair of Apple's earbuds. That teardown, published in late September, ended the way most earbud teardowns end: adhesive in every seam, a battery pressed flat against a wall, and a verdict that argues for buying a case instead. Seven weeks later the same team shipped a very different artifact. Valve's Steam Machine, the Linux gaming box that launched in June, now carries 18 official repair guides covering the fan, the RAM, the SSD, the power supply, and a motherboard swap that runs past fifty steps. Placing those two documents beside each other explains more about the state of repair in November 2026 than any policy paper I read this year.

An 18-Guide Suite Is a Design Statement

Valve's machine isn't a charity project. The base model sells for $1,049 with a 512GB SSD and the 2TB version for $1,349, built around a semi-custom AMD Zen 4 CPU and an RDNA-3 GPU. What matters here is that the memory and the storage are removable assemblies rather than soldered to the board. The guides weren't reverse-engineered from a retail unit either. They were written from Valve's own service instructions, which means the company chose to publish the route into its own hardware.

Modularity of that kind is an economic decision before it's a moral one. Sockets, connectors and removable modules add cost to the bill of materials and force industrial designers to surrender volume. Manufacturers accept that cost when they expect buyers to keep a machine for six or seven years, or when the alternative looks worse. A console with a soldered processor and replaceable everything else is also an admission about the one component nobody can service, and about how much of the machine's value sits inside that component.

The suite is honest about difficulty, which is the part most manufacturer documentation gets wrong. The SSD swap is short. The motherboard procedure is not, and the guide says so with a photograph of every step. Documentation like that is worth more than a badge, because it tells a technician whether to attempt the job before the machine is in pieces on the bench.

The Manual Isn't the Expensive Half

A repair needs three things, and they aren't equally expensive. Access comes first, and it's the screws-versus-glue argument that has defined this fight for a decade. Instruction comes second, and its marginal cost collapses to nothing once somebody has photographed a teardown, which is why iFixit's library has grown to roughly 63,700 free guides and why training a model on that archive is far cheaper than writing it. Inventory comes last, and it's the one that behaves like a physical constraint rather than a publishing problem.

Counting part numbers makes the point. A phone sold in five colours across forty countries can spawn hundreds of distinct order codes, each needing a warehouse slot, a forecast, a price and a returns path. Keeping a fan on a shelf for the seven years a buyer might need it requires somebody to bet on demand for a low-value module in a specific country in a specific year. Retailers have no interest in that bet. Most manufacturers would rather not either, which is part of why iFixit runs a parts store alongside its wiki.

Handing over documentation costs a manufacturer almost nothing and buys a surprising amount of goodwill. Nobody's margin depends on a PDF. Building a global spare parts channel costs real money and competes directly with the sale of the next device. That asymmetry explains why guide announcements arrive easily while clear statements about parts pricing and multi-year availability are rare, and why the Steam Machine's guides now exist while the question of buying a replacement fan at list price in year four still depends on a warehouse.

Europe Wrote Parts Into Law. America Is Writing a Rubric.

Europe legislated the boring half. Repairability scores have been required at the point of sale for smartphones in the EU for a year now, which means a shopper comparing two handsets in a store in Lyon sees a number a regulator defined rather than one a marketing department drafted. When a score is mandatory, the design argument happens in engineering meetings months before launch, because somebody has to hit the number.

None of that is happening in the United States. iFixit's answer has been to work with the National Sanitation Foundation on a voluntary scoring standard, deliberately separate from its own teardown scores. The public draft went live this year and closed for comment on October 9. Phase one adapts the EU's Joint Research Centre smartphone methodology for American conditions. Phase two builds a general framework on EN 45554 and the developing IEC 63683-2, standards that describe how to measure repairability rather than how to score a single product.

A voluntary standard isn't regulation, and iFixit says so plainly. The leverage it offers is procurement. Schools, hospitals and corporate fleet buyers purchase in bulk from spec sheets, and a spec sheet needs a comparable number. If the NSF standard turns up in tender language, it shapes hundreds of thousands of units without a single vote in Congress.

Fifty-Five Steps Is Its Own Kind of Warning

The score hides the cost of the path it certifies. Take the motherboard procedure and price it in time. Fifty-five steps at even thirty seconds each is 27 minutes of pure hand movement, before diagnosis, before testing, before reassembly risk, and before anyone buys the board. A careful technician will take considerably longer than that sum suggests.

The board itself is a large share of a $1,049 machine, because the processor is soldered to it and there's no cheap donor part to harvest. Put the labour and the board together and the honest answer for a failed mainboard is often that the console becomes a parts unit. That's a rational outcome for the owner and a terrible one for the idea that repairability means longevity.

None of this is a knock on the guide, which does exactly what documentation should do. The gap sits in the metric. A repairability score records whether a path exists, and the Steam Machine passes that test comfortably. What the score can't tell you is whether walking the path costs less than replacing the device, which is the only question a person with a dead console asks before doing anything else.

FixBot Learned the Library, Not the Warehouse

IFixit's other significant move pointed at the instruction layer instead. FixBot, launched last December, is an assistant trained on the company's own corpus: more than 125,000 repair guides, forum threads and manufacturer service PDFs spanning roughly 72,000 products. It accepts photographs and voice input, and it shows the schematic it's reading, which is the right design choice in a domain where a confidently wrong torque spec can crack a housing. The free tier is generous, with a paid plan around $4.99 a month.

The useful thing about FixBot is what it removes. It removes search, which was genuinely the last part of repair that scaled badly, since matching a symptom to a procedure across an archive that large used to be a skill in itself. It doesn't remove the friction that matters on a Tuesday night with a stripped screw, and it can't post you a clip. Reading it as an agent with a hard boundary rather than as a chatbot clarifies the limit, in the same way that the constraints on agentic workflows become obvious the moment a task requires touching a physical object.

Which leaves the parts question exactly where it was. An assistant that identifies a failed component with high confidence makes missing inventory more visible rather than less. Every accurate diagnosis ends at the same door.

The Variable No Scorecard Captures

Every instrument in that table shares one blind spot. They evaluate a device once, on a bench, with a technician's tools and a technician's competence, at the moment of launch. None of them samples the product again in year four, when the warranty has expired, the module is out of production, and the only supplier is a grey-market seller with a storefront and a shipping estimate.

A better rubric won't close that gap on its own. What would close it is a second measurement taken years after launch: whether a replacement part can still be bought, at what price relative to the device, and by whom. Nobody in the standards process has agreed on who collects that data, and the manufacturers holding the inventory have little reason to volunteer it. The draft now sitting with NSF scores a phone in a lab. It doesn't call the same phone's owner in 2030.

The open problem is a measurement problem, and it's the interesting one. A scorecard built on design reviews tells you how a device was built. The number that decides whether it survives to year six lives in a distributor's spreadsheet, and until a standard samples that spreadsheet, repairability will keep describing intentions rather than the machines people actually own.