Newsroom

News & Engineering NotesMilestones, partnerships, and dispatches from the atomic frontier.

Engineering BlogOctober 2026

Sophon Puts 330 GB of Memory on the Chip: 100× Faster Than HBM, a Fifth the Heat, 44 °C While Generating Text

An HBM GPU has two thermal problems where Sophon, our chip design, has one. Its memory is a 775 µm stack of bonded DRAM dies standing beside a compute die of about 1,800 W, and DRAM must refresh twice as often above 85 °C, so the memory sets the coolant limit. Sophon's 330 GB sits in 22.4 µm of the same die as its logic: every tier is within 1.3 K of the base at 1,026 W, refreshing it costs 0.08 W at 25 °C and about 4 W at 85 °C, and the die runs at 44 °C on a standard cold plate while it writes. These are modeled figures from the whitepaper; here is the arithmetic.

Engineering BlogOctober 2026

Our Case for Sophon Is Speed: What a Megawatt Would Earn Against a GPU Cloud, and When the Premium Runs Out

By PhantaField Architecture Team

In our yardstick, a megawatt of GPU cloud rented mostly on six-year contracts earns about $19.4 million a year at 2026–27 prices. If a token's price rose in step with how fast it arrives, a megawatt of Sophon, our chip design, would earn about 65 times that at 2026 prices, and falling token prices would wear the premium away around 2038. That is one pricing rule, not a forecast, so we also set out the evidence against it and what would have to happen before anyone counts on it.

Engineering BlogOctober 2026

What PhantaField Offers Terafab's Space Chip: Density Without EUV, Memory Without HBM or CoWoS, Fewer Bit Flips in Orbit, and Less Mass to Launch

By PhantaField Architecture Team

In March 2026 Elon Musk said every fab on Earth adds up to about 2% of the compute Terafab needs, most of it bound for orbit, on chips that tolerate radiation and run hot. We take that goal at face value and ask what a fab, a chip and a package would have to be. Four answers follow, each with its catch: density without EUV, memory without HBM or CoWoS packaging, a memory cell so thin a cosmic ray has to graze it to flip it, and a package that cuts a gigawatt's compute hardware from 10,384 tonnes to 363 with the radiator unchanged.

Engineering BlogOctober 2026

Scale Compute With Technology, Not Electricity: The Barrow Scale and AI's Gigawatts

By PhantaField Architecture Team

Since ENIAC, the energy cost of one computation has fallen about twelve orders of magnitude, while the power computing draws rose between five and six. AI's hardware has lately grown mostly by drawing more power. Our case is that compute should scale by technology first and electricity second, and that John Barrow's inward ruler measures that progress where Kardashev's outward one cannot.

Company NewsSeptember 2026

PhantaField Releases White Matter: the Packaging That Puts a Gigawatt of AI Compute in One Room

White Matter is a sealed brick of forty liquid-cooled silicon interposer tiers — 404 kW in 1.76 litres, about 2,900 times the density of a rack. Bricks join to each other by passive fiber, so a gigawatt site becomes one coherent machine about seven metres across. The whitepaper and the product page are live today.

Engineering BlogAugust 2026

Moore's Law Has a Second Axis: How PhantaField Reopened Scaling by Growing Up Instead of Shrinking Down

By PhantaField Architecture Team

Moore's observation was never about making transistors smaller — it was about integrating more of them at falling cost. For sixty years the industry had one way to do that: shrink the feature. Low-temperature 2D-semiconductor growth opens a second way: multiply the layers. Sixty-four tiers of mature 28 nm silicon reach the effective density of a 3.5 nm-class die, without EUV, and the two axes compound.

Engineering BlogAugust 2026

Jeff Dean Named the 1,000× Energy Penalty in AI Inference. PhantaField Designed the Chip That Deletes It

By PhantaField Architecture Team

In a Y Combinator interview, Google's Chief Scientist said moving data in from HBM costs ~1,000× the energy of computing on it — and that batching exists only to amortize that penalty, at the price of latency. Sophon's architecture collapses the factor to ~3×, and the consequences are exactly the ones Dean asked for.

Company NewsJuly 2026

SpaceX Is Taking AI to Orbit — PhantaField May Have the Only Chip That Survives the Trip

As SpaceX builds TeraFab — its own chip fab — and pulls AI compute toward orbit, the physics of space changes what qualifies as an accelerator. Radiation and power — not FLOPS — become the binding constraints, and both point the same direction.

Company NewsJuly 2026

Beyond HBM: How PhantaField Could Reshape the AI Memory Market

Sophon does not try to build a faster HBM stack. By placing 330 GB of memory inside a monolithic-3D compute fabric, PhantaField is challenging the assumption that every advanced AI accelerator needs HBM at all.

Company NewsJune 2026

PhantaField Introduces Sophon: Compute-in-Memory ASIC for the Post-HBM Era

The PFG-1 Sophon ASIC solution launches — 330 GB of DRAM inside the compute array, 2.1 PB/s of weight bandwidth, one die that trains and serves an 80B model.

Engineering BlogJune 2026

The Memory Wall Is a Distance Problem

By PhantaField Architecture Team

Fetching a weight across a package costs more than the math it feeds. The fix isn't faster memory — it's shorter wires.

Engineering BlogJune 2026

What Is the Memory Wall in AI? A Plain-English Guide

By PhantaField

The memory wall is the growing gap between how fast a chip can compute and how fast it can move data from memory. Here is what it is, why AI made it worse, and how it gets solved.

Engineering BlogMay 2026

HBM Explained: Why AI Memory Is Scarce, Expensive, and What Comes Next

By PhantaField

High Bandwidth Memory is the bottleneck of the AI buildout. Here is what HBM is, why it is in short supply in 2026, and the architectural alternatives.

Engineering BlogApril 2026

Compute-in-Memory (CIM), Explained: Doing Math Where the Data Lives

By PhantaField

Compute-in-memory performs calculations inside the memory that stores the data, eliminating data movement. Here is how it works, and how analog and digital CIM differ.

Engineering BlogMarch 2026

Monolithic 3D vs Chiplets vs 2.5D: A Guide to Chip Stacking

By PhantaField

2.5D, 3D chiplets, and monolithic 3D all stack silicon — but at very different connection densities. Here is how they compare and why the distinction matters for AI.

Engineering BlogFebruary 2026

Tokens per Watt: The Real Metric for AI Inference Efficiency

By PhantaField

Peak TFLOPS/W flatters chips that are starved for data. For real inference, energy per token — tokens per watt — is the number that decides datacenter cost.

Engineering BlogJanuary 2026

From 200 mm to 300 mm: Scaling a Plasma-Photon Reactor

By PhantaField Process Team

The physics worked on 8 inches. Production meant making it boring on 12 — four matched chambers, 60 wafers an hour, no drama.

Company NewsDecember 2025

PhantaField Debuts 12-Inch NanoGalaxy PPMOCVD at IEDM 2025

Two years after the 8-inch demonstrator, the production-scale 300 mm platform arrives: four chambers, 60 wafers per hour, BEOL-compatible growth.

Company NewsMay 2025

PhantaField Closes Angel Round

Eighteen months after Prof. Chenming Hu's first check, the angel round closes — funding the 300 mm platform program and the Sophon architecture team.

Engineering BlogJanuary 2024

Growing Crystals on Finished Chips: The Plasma-Photon Recipe

By PhantaField Process Team

Conventional MOCVD needs ~750 °C — hot enough to destroy a finished wafer. Here's how plasma and photons split that job between them.

Company NewsDecember 11, 2023

PhantaField Debuts 8-Inch MOCVD System at IEDM 2023

First public demonstration of plasma-photon MOCVD: wafer-scale 2D TMD monolayers grown on 200 mm wafers at BEOL-compatible temperatures.

Company NewsNovember 2, 2023

PhantaField Receives First Angel Investment from Prof. Chenming Hu

The father of the FinFET backs PhantaField's founding thesis: 2D semiconductors stacked into monolithic 3D are the path beyond silicon scaling.