Daniel Okafor

Covers AI accelerators from every vendor, the chip supply chain, GPU rental prices and export controls on AI chips. Works from chipmakers' product pages, datasheets, press releases and earnings materials, cloud providers' published price lists and official regulatory notices. A pen name of The Model Press research desk.

Latest

Dark navy cover reading "Google TPU 8i Max and TPU 8tx" with abstract square tiles labelled 8T, 8I, 8TX and 8I MAX, two tiles grouped for 8i Max
AI chip news ·

Google TPU 8i Max and TPU 8tx, announced November 9, 2026: what Google published

TPU 8i Max is two TPU 8i chips on one package; TPU 8tx pairs TPU 8t silicon with AMD EPYC hosts.

Dark navy cover with the title "AMD's Taalas deal: the $2.9B price" and a stacked bar splitting the deal value into about $1.5 billion not contingent and up to $1.4 billion contingent, from AMD's 10-Q of November 4, 2026
AI chip explainers ·

AMD Taalas acquisition: the $2.9 billion price and what Taalas builds

AMD's 10-Q puts the Taalas deal at about $2.9 billion, up to $1.4 billion contingent, plus the HC1 chip's specs.

Typographic cover: "SambaNova SN50 in Supermicro racks" beside a grid of outlined squares labelled prefill, a dashed arrow, and a row of cyan bars labelled decode.
AI chip news ·

SambaNova SN50 heads to Supermicro inference racks on November 3, 2026

SambaNova SN50 specs, and Supermicro's November 3, 2026 plan for 10,000 SN50 chips paired with Nvidia GPUs.

Dark navy cover reading "AI inference vs training compute", with an abstract cyan block grid labelled train feeding a stream of small tiles labelled serve.
AI chip explainers ·

AI inference vs training compute: what's the difference

Training builds a model and inference runs it, and the two stress AI hardware in different ways.

Typographic cover: "AMD Helios racks running at Meta" beside an outlined rack of nine slots, the top three filled in cyan, labelled rack, test, deploy.
AI chip news ·

AMD Helios racks running at Meta, AMD says on October 27, 2026

AMD said on October 27, 2026 that its first Helios racks are installed and in testing at Meta, ahead of a wider rollout.

The Model Press explainer cover for Huawei Ascend AI chips, with abstract teal processor blocks on a pale blue background.
AI chip explainers ·

Huawei Ascend: China's AI chips explained

Huawei Ascend AI chips from the 910 series to the 950, 960, 970 and 980, as Huawei described them at Huawei Connect.

Cover for The Model Press explainer: Intel Gaudi, specs and where it stands, with abstract teal blocks on a light blue background
AI chip explainers ·

Intel Gaudi explained: specs and where it stands

Intel Gaudi explained: Gaudi 3 specs from Intel's own paper, where to get it, and Intel's next AI chip.

Typographic cover: "HBM suppliers: SK hynix, Samsung and Micron in the HBM race" beside stacked teal blocks of different sizes.
AI chip explainers ·

HBM suppliers: SK hynix, Samsung and Micron in the HBM race

SK hynix, Samsung and Micron all say they ship HBM4 in volume as of October 2026; here is where each supplier stands.

Cover for the TSMC and AI chips explainer: how TSMC fabricates and packages AI accelerators
AI chip explainers ·

TSMC and AI chips: why so many accelerators run through it

TSMC makes and packages the silicon behind Nvidia GPUs, Broadcom XPUs and Intel Gaudi 3: nodes, CoWoS and Arizona.

Cover illustration: teal block shapes on a pale background beside the title "Broadcom and custom AI chips (XPUs): who designs them", labelled Explainer · Chips
AI chip explainers ·

Broadcom AI chips: custom XPUs and who designs them

Broadcom's AI chips are custom accelerators (XPUs) built for Google, Meta, OpenAI and others.

Typographic cover: "Vera Rubin on Google Cloud A5X" beside four outlined site blocks of GPU cells joined by dashed links, labelled 80K per site and 960K multi-site.
AI chip news ·

Google Cloud A5X with Vera Rubin NVL72 announced October 21, 2026

Google Cloud A5X instances on Vera Rubin NVL72, announced October 21, 2026, reach general availability next month.

Typographic cover: "vLLM and SGLang run natively on TPU" beside a stack diagram with vLLM and SGLang boxes above PyTorch and a solid TorchTPU layer.
AI chip news ·

vLLM and SGLang on TPU: Google's PyTorch engines, October 20, 2026

Google's new vLLM and SGLang engines run natively on TPU through TorchTPU, launched October 20, 2026.

Dark navy cover reading "Explainer · Chips: AI chip export controls", with abstract bars of rising height crossing two dashed threshold lines labelled TPP
AI chip explainers ·

AI chip export controls: what the US rules restrict

US AI chip export rules as of November 5, 2026: TPP limits worked through, the China and UAE rules, the affiliates rule.

AI chip explainers ·

Nvidia's data-center GPU roadmap, Hopper to Rubin

Nvidia went from Hopper through Blackwell and Blackwell Ultra to Vera Rubin in four years; here is each generation.

AI chip explainers ·

Groq AI and its LPU: what the inference chip does

Groq built the LPU, an inference chip that keeps model weights in on-chip SRAM, and runs it as the GroqCloud service.

AI chip explainers ·

Cerebras Systems: the wafer-scale chip and its business

Cerebras Systems builds AI processors from a whole silicon wafer and sells both systems and inference cloud access.

AI chip explainers ·

AWS Trainium: what it is and which models train on it

AWS Trainium is Amazon's own AI chip, rented through EC2 Trn2 and Trn3 instances and UltraServers.

Abstract network fabric linking three clusters of AI accelerators with blue, green and amber data lanes
AI chip news ·

Ultra Ethernet 1.1 adds scale-up links and collectives

Ultra Ethernet 1.1 adds a local accelerator profile, in-network collectives and new congestion controls.

AI chip explainers ·

Google TPU: generations and who can use them

Google's TPUs run Gemini and are rented through Google Cloud, from v5e to Ironwood and the announced TPU 8t and 8i.

KV cache diagram showing token blocks reused from memory to generate the next token.
AI model explainers ·

KV cache explained: how it speeds up AI inference

A KV cache reuses attention keys and values during generation, saving compute while using more accelerator memory.

Typographic cover: "Azure Maia 300: 96 accelerators in one rack" beside a grid of 96 outlined cells arranged as 24 rows of four, with a few highlighted in cyan.
AI chip news ·

Microsoft unveils Azure Maia 300 on October 12, 2026: 96-chip racks

Microsoft unveiled Azure Maia 300 on October 12, 2026: 288 GB of HBM4 per chip, 96 chips per 167 kW rack.

Typographic cover: "HBM memory, HBM2E to HBM4E, stack by stack" beside four cyan memory stacks, two 8-high and two 12-high.
AI chip explainers ·

HBM memory explained: HBM2E, HBM3, HBM3E and HBM4 compared

HBM2E, HBM3, HBM3E, HBM4 and HBM4E compared by width, pin speed, bandwidth and capacity per stack.

AI chip explainers ·

AMD Instinct MI355X: specs and how it compares

The AMD Instinct MI355X is AMD's CDNA 4 accelerator with 288 GB of HBM3E at 8 TB/s and 1,400 W board power.

NVIDIA RTX Spark preorder headline on a dark grid with a teal circle and ring.
AI chip news ·

NVIDIA Opens RTX Spark Preorders for Windows AI PCs

Laptops are due October 16 and compact desktops in November; DGX Station for Windows remains a preview.

The Model Press

What are you looking for?

Search by headline, topic or keyword.