🗞️ AI & GPU Industry Weekly Recap: February 2–8, 2026

🔑 Key Highlights

AMD shatters records with Q4 2025 revenue of $10.3 billion (up 34% YoY) and full-year 2025 revenue of $34.6 billion, driven by explosive Data Center and Client segment growth
AMD’s Data Center segment hits a record $5.4 billion in Q4 alone, fueled by EPYC CPU adoption and AMD Instinct GPU shipments — with full-year Data Center revenue reaching $16.6 billion
AMD launches Micro-World, its first fully open-source interactive world model for action-controlled video generation, built on the Wan 2.1 architecture and trained/deployed on AMD Instinct MI325X GPUs
NVIDIA DLSS integration with Blender stalls over licensing concerns, as the Blender team hesitates to ship proprietary DLSS binaries alongside the open-source renderer — keeping the door open for Intel Open Image Denoise 3 as a cross-vendor alternative
NVIDIA CEO Jensen Huang and Dassault Systèmes announce their largest-ever collaboration, framing AI as infrastructure and pushing “industry world models” powered by NVIDIA Omniverse, BioNeMo, and CUDA-X libraries

🤖 AI & Machine Learning

AMD Micro-World: Open-Source World Modeling

AMD’s AI research team (AMD-AGI) published Micro-World, the company’s first fully open-source world model targeting interactive video generation. Key technical highlights:

Built on Wan 2.1 (both 1.3B T2V and 14B I2V variants)
Trained on 6,000 Minecraft gameplay clips (81 frames each) with keyboard + mouse action annotations and text captions generated via miniCPM
Uses a two-stage training paradigm: first, a game-style LoRA is trained; second, an action processing module is trained separately (with base model frozen) for open-domain transferability
Action injection explored via ControlNet (T2W variant) and Adaptive Layer Normalization / adaLN (I2W variant)
Fully open-sourced: model weights, training/inference code, and curated dataset released publicly at github.com/AMD-AGI/Micro-World
Training and inference conducted on AMD Instinct MI325X GPUs with ROCm, using Accelerate, Diffusers, and DeepSpeed

GP-MoLFormer on AMD Instinct MI300X

AMD’s ROCm team published a detailed blog on running IBM’s GP-MoLFormer — a generative molecular foundation model — on AMD Instinct MI300X accelerators. The model, trained on over 1 billion canonical SMILES strings, supports:

Unconditional molecular generation
Scaffold-constrained conditional generation
Property-guided design via pair-tuning
Evaluated on the MOSES benchmark (Frag, Scaf, SNN, IntDiv, FCD) and MoLFormer-based metrics (FMD, DNN)

The blog positions the MI300X as a competitive platform for Scientific AI, complementing earlier work on the MatterGen crystal generation model, also run on AMD hardware.

SparK: KV Cache Compression Research

AMD’s ROCm blog also featured updates on SparK (from arXiv:2508.15212), a technique addressing the KV cache bottleneck in long-context LLM inference. SparK applies unstructured channel pruning during prefill and recovers pruned channels during attention score computation — a distinct approach from eviction-based or structured pruning methods.

NVIDIA + Dassault Systèmes: Industrial AI World Models

Jensen Huang appeared at 3DEXPERIENCE World in Houston to announce a landmark partnership with Dassault Systèmes CEO Pascal Daloz. The collaboration targets “industry world models” — physics-grounded AI systems for engineering simulation. Key components:

NVIDIA BioNeMo + BIOVIA for molecular/materials discovery
SIMULIA AI Virtual Twin Physics using NVIDIA CUDA-X libraries
NVIDIA Omniverse + DELMIA for software-defined factory simulation
NVIDIA Nemotron open models powering “Virtual Companions” within the 3DEXPERIENCE platform
Dassault’s OUTSCALE sovereign cloud deploying NVIDIA-powered AI factories across three continents

⚡ GPU & Hardware

AMD Q4 2025 Hardware Highlights

AMD’s earnings call surfaced several hardware milestones:

Ryzen 7 9850X3D launched as AMD’s fastest gaming processor, combining Zen 5 architecture with AMD 3D V-Cache
AMD Instinct MI440X announced for enterprise AI at CES 2026
AMD Helios rack-scale platform previewed — a blueprint for yotta-scale AI infrastructure; HPE committed as an early adopter, building the Herder supercomputer with next-gen Instinct MI430X GPUs and EPYC “Venice” CPUs
AMD FSR “Redstone” suite unveiled: ML-based upscaling, Frame Generation, Ray Regeneration, and Radiance Caching for Radeon GPUs
AMD Instinct MI308 China export saga continues: ~$390M in MI308 revenue to China in Q4, with ~$360M inventory reserve reversal. Q1 2026 guidance includes ~$100M in additional MI308 China sales

AMD Kintex UltraScale+ Gen 2 FPGAs

AMD expanded its FPGA lineup with the Kintex UltraScale+ Gen 2, targeting mid-range, performance-critical applications in medical and industrial markets. It supports 4K/8K media workflows, advanced imaging, and offers a migration path from existing Spartan UltraScale+ FPGAs. Pre-production sampling expected Q4 2026; volume production slated for H1 2027.

NVIDIA RTX 50 Series Supply Chaos

The NVIDIA GeForce RTX 50 “Blackwell” series supply situation remains dire. GPU prices are significantly above MSRP at major retailers, with near-zero inventory across U.S. channels. Ironically, some units are appearing in Walmart clearance aisles at massive discounts — one shopper reportedly acquired a PNY RTX 5080 (MSRP: $999) for $289.49 (71% below MSRP), while RTX 5080s sell for as high as £1,399 (~$1,918) in the UK. The shortage is attributed to GDDR7 memory supply constraints and intense AI-driven demand.

NVIDIA DLSS in Blender: Stalled Over Licensing

NVIDIA submitted a pull request to integrate DLSS Ray Reconstruction into Blender’s Cycles renderer for viewport denoising. The Blender render team acknowledged impressive quality but raised concerns:

The integration requires shipping NVIDIA proprietary binaries (nvngx_dlssd.dll / libnvidia-ngx-dlssd.so) alongside the open-source application
The Blender team prefers the implementation to be embedded in the NVIDIA driver itself
Intel Open Image Denoise 3 (expected H2 2026) is cited as a preferred open-source, cross-vendor alternative
Minimum required NVIDIA driver: 590+

Loongson 3B6000 vs. AMD Zen 5

Phoronix published benchmarks of China’s Loongson 3B6000 (12-core LoongArch, SMT2, DDR4 ECC) versus AMD Zen 5 (Ryzen), Intel Arrow Lake, and Raspberry Pi 5. The review marks the first LoongArch hardware benchmarks on Phoronix and highlights the growing importance of domestic Chinese CPU alternatives, though performance versus leading AMD and Intel desktop platforms remains the key question.

🏭 Industry & Market

AMD FY2025: A Defining Year

AMD’s full-year 2025 results represent a fundamental shift in the company’s scale and profitability:

Metric	FY2025	FY2024	YoY
Revenue	$34.6B	$25.8B	+34%
Non-GAAP Operating Income	$7.8B	$6.1B	+27%
Non-GAAP EPS	$4.17	$3.31	+26%
Data Center Revenue	$16.6B	~$12.6B	+32%
Client + Gaming Revenue	$14.6B	~$9.7B	+51%

Notable callouts:

Client business revenue hit a record $10.6B (+51% YoY), driven by Ryzen market share gains
Gaming revenue reached $3.9B (+51% YoY), boosted by semi-custom wins and Radeon GPU demand
Q1 2026 guidance: ~$9.8B revenue (+32% YoY), non-GAAP gross margin ~55%
ZT Systems’ manufacturing business was divested in Q4 2025

Strategic Partnerships and Joint Ventures

AMD + Cisco + HUMAIN announced a joint venture targeting 1 GW of AI infrastructure by 2030
AMD + Tata Consultancy Services partnered to co-develop enterprise AI solutions
AWS launched new instances powered by 5th Gen AMD EPYC CPUs, claiming highest x86 performance in the AWS cloud
Zyphra’s ZAYA1 became the first large-scale MoE model trained entirely on AMD Instinct MI300X GPUs, using AMD Pensando networking and ROCm software

NVIDIA’s Industrial AI Push

NVIDIA is positioning itself as the backbone of industrial simulation, partnering with Dassault Systèmes (400,000+ customers, 45M+ users of 3DEXPERIENCE) to embed NVIDIA AI across engineering, manufacturing, and biology workflows. This represents NVIDIA’s clearest articulation yet that its TAM extends well beyond training clusters into industrial digital twins and physical AI.

🛠️ Developer Ecosystem

ROCm Ecosystem Expansion

AMD’s ROCm team shipped multiple developer-facing blog posts this week, covering:

GP-MoLFormer molecular generation on MI300X using rocm/pytorch:rocm7.0_ubuntu22.04_py3.10_pytorch_release_2.7.1 Docker images
Micro-World training stack: PyTorch + Accelerate + Diffusers + DeepSpeed on MI325X GPUs
SparK KV cache optimization documentation corrections

The GP-MoLFormer setup notably falls back gracefully when the fast_transformers dependency is unavailable, using standard PyTorch transformer implementations — a sign of improving ROCm compatibility with third-party ML libraries.

Wan 2.1 as an Open Ecosystem Base Model

AMD’s choice of Wan 2.1 (both 1.3B and 14B variants) as the foundation for Micro-World signals growing developer momentum around this open-source video generation architecture. AMD’s contributions (weights, training code, dataset) are positioned to extend the Wan 2.1 ecosystem into action-controlled and interactive video domains.

NVIDIA Driver Ecosystem Friction

The Blender DLSS episode highlights a recurring tension: NVIDIA’s most compelling features (DLSS, OptiX) require proprietary driver components that conflict with open-source distribution norms. With Intel Open Image Denoise 3 targeting cross-vendor GPU support in H2 2026, there may be meaningful ecosystem pressure on NVIDIA to rethink its binary distribution model for creative tools.

📊 Key Takeaways

AMD’s Q4 and FY2025 results confirmed that its multi-year bet on EPYC CPUs and Instinct GPUs has matured into a genuine revenue engine — with $34.6 billion in annual revenue and a Data Center segment now generating over $16 billion per year, AMD is firmly a Tier-1 AI infrastructure vendor and no longer a distant challenger to NVIDIA. Meanwhile, AMD’s ROCm team is aggressively expanding its open-source software credentials this week, shipping Micro-World (its first open-source world model) and GP-MoLFormer support on MI300X, signaling that AMD’s competitive strategy in the developer ecosystem increasingly mirrors its hardware ambitions. The broader market backdrop — NVIDIA RTX 50 series GPU shortages driving absurd retail price premiums, DLSS licensing friction in Blender, and NVIDIA doubling down on industrial AI with Dassault Systèmes — underscores that the GPU industry in early 2026 is simultaneously supply-constrained, software-competitive, and rapidly expanding its definition of what “AI infrastructure” means.

News Weekly: 2026-02-02–2026-02-08