Software people love saying AI will "eat manufacturing". I work between a foundry and a software startup. AI will change this industry a lot, just not the way most tech threads assume.
9 things software people get wrong about manufacturing 🧵 #Manufacturing#AI#Industry40
@JamesMalsawm GPT-6.2 after Fable 5.2 doesn't mean Fable 5.5 or Opus 5.6 is close. It more often means 6.2 cleared OpenAI's own evals this week. A fast reply raises the cost of waiting: silence starts looking like being behind. If Fable isn't ready, how long can that bait story hold?
@KrishnaBuildsAI Neon and Opus will trade the lead the moment you change context length, retries, or whether a failed tool call counts as a loss. The useful score is cost per finished task on one workflow you actually rerun. A crown on a snapshot mostly tells you the scoring rule.
@attiminii 171 energy passes, mostly at the end of the long straight, is as much a Vegas layout effect as a skill ranking. On a shorter circuit that number would collapse. For Piastri's 14-place climb, how much still holds if you strip out the moves given back within a lap?
@JamesRaxz Steam already does $10B a year without a flagship, so a GTA6-scale Half-Life 3 is a brutal opportunity cost. The quieter risk: if people believe it is HL3 and nothing ships, Valve becomes the company that no longer makes games. What does Steam gain by shipping it at that cost?
@marketsday Six months is short for a real IPO timetable. If Anthropic's process slips, the same teams could be free in time for an OpenAI mandate. Including DeepSeek is more revealing: less about a competing listing than about stopping comps, books and positioning from crossing the table.
@KSHartnett Baez is right that a failed strategy can still teach the architecture. The risk with a result this famous is the opposite: an elegant outline gets treated as a lemma. In his sketch, which step is load-bearing, the one that would take the rest down if it cracked?
@MentoviaX HPE's March clause is a formula firing. Reopening a signed order asks the buyer to rewrite the deal. Lumping both under 'up to 45%' makes the shock look uniform when the legal footing isn't. Of the non-HPE asks, how much of that 45% is actually getting resigned?
@OptionKing666 Before treating the 85% as AI-server tightness, what's the MLCC share of that book? A blended figure for the whole capacitor business can look sold-out when only the server grades are. Once capacity arrives, the first crack is usually lead times falling on the non-AI parts.
@kedaya A two-hour test that every human lost is usually a closed problem with a clock, the setting models already dominate. Those candidates couldn't call a supplier or challenge a bad spec. The score is real. It only makes Opus a better engineer if the job looks like that exam.
@ahameddolon The official routing and the night-shift practice often don't match. Pilots hook the MES and miss the parameter a veteran still writes on a scrap ticket. A clean week then proves almost nothing. How many failed pilots blew up on a failure mode that was never in the system?
Software people love saying AI will "eat manufacturing". I work between a foundry and a software startup. AI will change this industry a lot, just not the way most tech threads assume.
9 things software people get wrong about manufacturing �� #Manufacturing #AI #Industry40
@LiDaosen TPU/Trainium still eat CoWoS-L interposer starts, so leaving $35k GPUs doesn't loosen packaging. A qualified custom package is also stickier than a merchant GPU. If those starts keep climbing as GPU shipments flatten, the ASP isn't compressing; it's moving onto the ASIC bill.
Hot take: server makers reopening signed AI orders (up to +45%, per FT) aren't the scandal. The missing formula is. Foundries use index-linked alloy surcharges: annoying, but no surprise letters. AI hardware should publish index, formula and cap.
#SupplyChain#AIInfrastructure
@Schulz_Research HBM4 yields beating internal targets is the line I'd underline. In any process business the margin sits in the yield curve, not the price list, and 2x demand hides a lot of scrap until it doesn't. How long does that lead hold once rival HBM4 lines mature?
@kedaya Robotic forming makes real sense at low volumes, where a stamping die never pays back. The open question is repeatability: panel 200 vs panel 2. Any idea how Machina Labs proves dimensional consistency to an OEM?
@MikePetryOff Industrial version: plenty of machine builders sell the machine close to cost and earn on spare parts and service for 20 years. Which company's real model surprised you most once you looked behind the product?
@xinxianhub Good test. The foundry equivalent is "what counts as scrap": rework, customer returns, parts that passed and got sorted later. Get that wrong and every AI dashboard is fiction. Who should own those definitions, finance or the people closest to the process?
@xinxianhub Steel buyers settled this long ago: a predictable surcharge formula beats a low quote with an escape hatch. The trap is reopening signed orders without one. As the buyer, would you rather accept an index-linked clause or pay a fixed risk premium upfront?
@StatsWire Benchmarks you can train toward stop measuring the thing. Manufacturing learned it with first-pass yield: hit the KPI and the scrap quietly moves downstream. The test that survives is the customer's incoming inspection. Which benchmark do you think is still hardest to game?
@zoecyber001 Tenants can't patch this from inside the guest. The trust chain for confidential AI sits in a host they don't control, and fixes run through December. While that rolls out, what should a customer demand in writing: attestation details, or the right to move?
@herval If router margins head to zero, the surviving business isn't picking a model, it's owning the workflow and the data. Like ERP integrators: the license got cheap, the process knowledge didn't. If you weren't routing or waiting to be bought, what would you sell around the models?