For two years the AI industry has been told the same thing: the next model needs more memory, and more memory means more HBM. But HBM is the one component the entire frontier is fighting over. Supply is locked up years in advance, prices have climbed faster than any logic node, and a single accelerator now spends more of its bill of materials on memory than on the silicon that does the math. PhantaField today introduced Sophon, an AI chip architecture that answers the memory wall by refusing to participate in that market at all.
Sophon grows its memory directly above its compute. The PFG-1 reference design stacks 80 tiers of 2D transition-metal-dichalcogenide transistors on a mature 28 nm silicon base — 40 logic tiers interleaved with 40 memory tiers — and embeds 330 GB of capacitor-less 2T0C DRAM inside a pure-digital compute-in-memory array. Because every weight sits a few hundred nanometers above the multiply-accumulate unit that consumes it, the die delivers 2.1 PB/s of in-tile weight bandwidth: roughly 100× an eight-stack HBM4 package, at a bill of materials near $7,400 — under a tenth of a 2026 HBM4 GPU module.
One die that trains and serves
Capacity is the half SRAM accelerators cannot match. Where wafer-scale and chiplet designs hold megabytes to tens of gigabytes on-chip and must shard a single model across hundreds of devices, Sophon fits an entire 80-billion-parameter model — weights, optimizer state, and activation headroom — on one die. The same silicon trains in BF16 and serves low-batch inference at the compute-bound rate, so a fleet can be elastically repartitioned between training and serving without touching the hardware. At batch one it decodes that 80B model at 14,438 tokens per second in FP8, at 25.8 millijoules per token — roughly 174× less energy than an HBM4 GPU spends at the same point.
The bottleneck doesn't get faster. It ceases to exist.
PhantaField
Three ways to build on it
Sophon reaches customers three ways: as silicon — packaged dies and accelerator boards for AI labs, clouds, and enterprises; as a managed training service — bring your data and model, and train on Sophon clusters; and as a hosted, OpenAI-compatible inference API, paid per token. One architecture, owned end to end — from the deposition tool that grows the 2D layers to the chip that serves the tokens.
The complete technical case — cell physics, the digital CIM tile model, thermal analysis, yield, and a full economic comparison against the 2026 HBM4 flagships — is published in the PFG-1 whitepaper. Partners can request an NDA-protected briefing through the contact form.