Cerebras Releases Tungsten to Program Its Wafer-Scale Chip

Cerebras Systems released Tungsten, a language and compiler for programming its Wafer-Scale Engine at a high level, in a blog post on 16 November, with Aviary, a set of applications and tutorials written in it. Co-founder Michael James wrote that last month a Cerebras researcher ran a plasma physics simulation across 98% of a Cerebras wafer. “A coding agent wrote every line of it”, he wrote.

Cerebras describes the wafer as a dataflow machine with 900,000 cores. 98% of that is about 882,000 cores, by our arithmetic; the post does not give the run’s exact core count. Both tools are live at labs.cerebras.ai, which labels itself a “research preview”.

Business Pill · Switching Costs

An evergreen Business Engineer explainer on switching costs: why it is hard to leave an AI supplier. It teaches the general idea only and does not discuss Cerebras.

The key insight: As we read it, the news is less the 1.6 PFLOP/s than who wrote the showcase code: Cerebras says a coding agent wrote every line of a plasma simulation that ran across 98% of a wafer, about 882,000 of its 900,000 cores by our arithmetic.

What Tungsten Replaces

Before Tungsten, Cerebras’ compilers exposed the wafer’s vector, stream and event hardware as intrinsics. “In practice, every non-scalar program was written close to assembly”, James wrote. Programmers configured descriptors, wired tasks to events and routed the fabric by hand.

Tungsten’s compiler “carries a full model of the architecture” and maps code onto access pattern registers, event-driven tasks, fabric calls and loop-bearing instructions, the post says. In the tutorial’s Jacobi stencil example, each line is a loop over a whole block and compiles to a single vector instruction, with no descriptor setup or task wiring.

Cerebras compares this with register allocation on CPUs. The compiler assigns access patterns, loop definitions, fabric sockets and microthreads with the whole program in view, and swaps them in and out as the program runs. “Because the compiler owns these decisions, restructuring a code later is an edit, not a rewrite”, the post says.

How code for Cerebras’ Wafer-Scale Engine was written before Tungsten and what its compiler now assigns,
How code for Cerebras’ Wafer-Scale Engine was written before Tungsten and what its compiler now assigns, with the agent-written plasma run’s figures as Cerebras reports them. About 882,000 cores is our arithmetic (98% of 900,000). Source: Cerebras blog and labs.cerebras.ai, 16 November 2026.

The Numbers Cerebras Gives

Cerebras says SNAP, a standard proxy for nuclear radiation transport codes, runs at 1.6 PFLOP/s on a single CS-4, “ahead of 10,800 GPUs on the El Capitan supercomputer”. It says HPCG, the benchmark used to rank supercomputers on real-world sparse workloads, reaches 1.04 PFLOP/s on a single system.

The post says Cerebras’ wafer-scale molecular dynamics runs 748 times faster than Frontier, one of the world’s largest supercomputers, citing a 2025 paper in the Journal of Chemical Physics. On some problems, it says, a wafer runs “hundreds of times faster than the largest supercomputers”.

The agent-written particle-in-cell run advanced 60 million particles at 7,200 time steps per second, or 430 billion particle updates per second on a single system, according to the post. These are Cerebras’ own figures; we did not reproduce any of them.

Built for Coding Agents

One researcher and one coding agent built the plasma code in about two weeks in a chat window, the post says. The agent worked from published algorithms, validated its reference against the open-source XOOPIC code and ran it on a problem modeled on the EMPIRE plasma code. It also wrote the tests, figures and documentation, and found and fixed its own bugs, James wrote.

The tutorial has 12 lessons and 21 topics, and Cerebras says every lesson and topic is built, run in a simulator and checked by its own test. The Labs SDK home page says the tutorial “is ready for coding agents”. James added a caveat: “Someone has to keep the agents on track”.

What Ships, and Where It Came From

Aviary lists 10 released applications, by our count: a heat-equation stencil, a Gray-Scott reaction-diffusion code, shallow-water equations, SNAP, HPCG, a distributed fast Fourier transform, two molecular dynamics codes, the particle-in-cell code and high-order hydrodynamics. The Gray-Scott tutorial code is about 50 lines of Tungsten. A design optimizer that couples an AI model to a physics simulation is listed as forthcoming.

Many of the applications grew out of work with Sandia, Lawrence Livermore and Los Alamos national laboratories under NNSA’s Advanced Memory Technology program, Cerebras says. The same Tungsten code runs on one wafer or across a cluster, and its global tsunami simulation ran on 64 wafers, the post says.

The Labs SDK comes as an archive whose programs run on x86-64 Linux with glibc 2.34 or newer, according to its setup page. The archive includes the tungsten compiler, a simulator and the tutorial with the applications.

Why a Compiler Matters to a Chip Company

As we read it, Tungsten targets the cost of writing code for the wafer rather than the chip itself. Nvidia’s position in AI accelerators is widely tied to CUDA, the software layer much GPU code is written for. James wrote that the old intrinsics style “also took deep expertise, and the result was hard to read, change or reuse”.

In our reading, two parts of the release go at that cost: a language whose code reads close to the textbook algorithm, and a reference complete enough for a coding agent to write a full application. Buyers already mix suppliers: six of eight AI firms mix chips, per our map of Nvidia’s allies and rivals. AMD made 12.4% of vLLM’s commits in a year, 2.6 times Nvidia’s, per our piece on AMD’s commits to vLLM.

Cerebras raised its 2026 outlook to $950 million last week, per our piece on its outlook. The 10 Aviary applications are scientific simulation codes rather than AI models, by our reading of the list.

The Structural Read

As we read the post, Tungsten turns resource choices that programmers used to make by hand (descriptors, task wiring, fabric routing) into compiler decisions, the way CPU compilers took over register allocation, which is the comparison Cerebras itself draws.

As we read the release, the agent example and the simulator-checked tutorial work toward the same goal: lowering the expertise a team needs before its first useful program runs on the wafer. Cerebras gives no adoption figures, so how far that goes is not shown.

Michael James, Cerebras co-founder, in the Tungsten announcement

“When a high-performance code no longer takes a specialist to write, more problems become worth attempting.”

Three Implications

FOR SCIENTIFIC COMPUTING TEAMS The Labs SDK runs on x86-64 Linux and includes a simulator, so a program can be built and checked before it reaches a wafer, per the setup page.

FOR CODING-AGENT USERS Cerebras says its plasma code came from one researcher and one coding agent in about two weeks; James still advises learning the language.

FOR READERS WEIGHING THE NUMBERS SNAP at 1.6 PFLOP/s, HPCG at 1.04 PFLOP/s and 430 billion particle updates per second are Cerebras’ own figures, not independent results.

The Business Engineer Lens

This story maps onto the Business Engineer framework Enterprise AI: From Software to Substrate.

The framework argues that “enterprise vendors don’t sell software anymore—they sell irreversibility”, and that “switching costs don’t come from learning curves or data migration headaches”.

As we read the release, Tungsten works on the learning-curve side of the cost of adopting a wafer: code that reads close to the textbook algorithm, and a reference an agent can learn from. Code written in Tungsten still targets Cerebras hardware, so in the framework’s terms the deeper test is how many of a lab’s other systems come to depend on it. The framework does not discuss Cerebras or Tungsten.

What Is Not Established

The pages we read do not state Tungsten’s licence terms or a price, or say whether access to the Labs SDK download is restricted. They do not name users of Tungsten outside Cerebras and its national-laboratory work. The benchmark figures are Cerebras’ own, and we did not reproduce them. We did not contact Cerebras for this piece.

This Week In The AI Supercycle

AI Lost Its Central Banker, the Lab Oligopoly, and the Future of Open Source

This week’s issue: the Nvidia leadership change, read as the AI build-out losing its central banker: who now finances the GPU landlords, backstops the data centres and sponsors open models, and why two labs are becoming the bank.

Read the full issue →

Business Engineer Framework

Enterprise AI: From Software to Substrate

The Business Engineer’s Embedding Mental Model: why switching costs come from integration depth, not learning curves.

Read the Map of AI →

The Bottom Line

Cerebras released Tungsten, a high-level language and compiler for its 900,000-core Wafer-Scale Engine, and Aviary, 10 applications and a tutorial written in it, on 16 November. Its lead example is a plasma code that a coding agent wrote in about two weeks and that ran across 98% of a wafer, about 882,000 cores by our arithmetic.

95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.

A note on sourcing. We read Cerebras Systems’ blog post introducing Tungsten and Aviary, by co-founder Michael James, in full, along with the labs.cerebras.ai home page and its Run the tutorials page, and the tutorial index. Performance figures are Cerebras’ own and were not independently reproduced. The core estimate and the application count are our arithmetic. We did not contact Cerebras for this piece.

Sources: Cerebras Systems blog: Introducing Tungsten and Aviary (Michael James), 16 November 2026 · Cerebras Labs SDK home page (research preview) · Cerebras Labs: Run the tutorials · Cerebras Labs: Tutorial

Scroll to Top
FourWeekMBA