I’ve not been able to find any solid intel on what actually happened to the wafers.
Thing is though- cutting wafers and packaging can be done at any fab so in theory they could find an older fab with excess capacity and pay them to do some of it.
Binning however requires some more specialized equipment, that from what I understand is more specific to the type of chips you’re making. This is where you test each individual chip- out of a wafer you’ll get some great ones, some good ones, some bad ones, and some that don’t work at all. Thus ‘binning’ is taking a stream of chips and sorting them into bins by their maximum speed, stability, etc.
You might find a fab that has the equipment to bin DRAM chips but not the means to manufacture those chips. Still it seems an odd use of resources to get into the chip production business when your core business has nothing to do with chips.
IMO, nothing (or, more accurately: everything) is odd in the AI industry. AI companies have had to buy compute resources from rivals. Hardware manufacturers can’t ramp up fast enough to meet demand. Data centers can’t be built fast enough. Energy production is a bottleneck. Three Mile Island is being resurrected. SpaceX now owns xAI, plans to buy Cursor, and is unironically considering building data centers in orbit. Software security vulnerabilities are being discovered/revealed at an alarming rate.
Honestly, OpenAI buying wafers seems kinda tame at this point 😅
I’ve not been able to find any solid intel on what actually happened to the wafers.
Thing is though- cutting wafers and packaging can be done at any fab so in theory they could find an older fab with excess capacity and pay them to do some of it.
Binning however requires some more specialized equipment, that from what I understand is more specific to the type of chips you’re making. This is where you test each individual chip- out of a wafer you’ll get some great ones, some good ones, some bad ones, and some that don’t work at all. Thus ‘binning’ is taking a stream of chips and sorting them into bins by their maximum speed, stability, etc.
You might find a fab that has the equipment to bin DRAM chips but not the means to manufacture those chips. Still it seems an odd use of resources to get into the chip production business when your core business has nothing to do with chips.
IMO, nothing (or, more accurately: everything) is odd in the AI industry. AI companies have had to buy compute resources from rivals. Hardware manufacturers can’t ramp up fast enough to meet demand. Data centers can’t be built fast enough. Energy production is a bottleneck. Three Mile Island is being resurrected. SpaceX now owns xAI, plans to buy Cursor, and is unironically considering building data centers in orbit. Software security vulnerabilities are being discovered/revealed at an alarming rate.
Honestly, OpenAI buying wafers seems kinda tame at this point 😅