AMD Instinct Coder 355 serves 85 developers with GLM-5.3 on MI355X

1 hour ago

9 min read

AMD has introduced the AMD Instinct Coder 355, the second generation of its turnkey on-premises platform for AI-assisted software development. The system was presented on November 9, 2026, in a post on the AMD blog signed by Kumaran Siva, corporate vice president for Enterprise AI, and it combines a Supermicro server with eight AMD Instinct MI355X GPUs, two AMD EPYC 9575F processors, Spectro Cloud's PaletteAI Inference Launchpad software and the open-weight GLM-5.3 coding model running locally.

AMD rates one node for up to 85 developers, with up to 50 concurrent users. The previous configuration, the Instinct Coder 325 built on MI325X GPUs, is listed on AMD's product page at up to 50 developers and 30 concurrent users per node, so the new generation raises the rated team size by 70% on the same eight-GPU footprint.

The platform is designed for companies that want coding agents such as Claude Code, Cursor and Visual Studio Code to run against a model hosted inside their own data center by default, with requests escalated to Anthropic Claude, OpenAI GPT or Google Gemini only when a policy allows it. AMD claims up to 75% lower total cost of ownership and up to 98% of frontier benchmark performance, and both figures rest on footnotes that describe a modeled scenario and a single benchmark configuration. No price or shipping date for the platform has been published, and AMD directs buyers to contact AMD, Supermicro or Spectro Cloud representatives.

··········

AMD INSTINCT CODER 355 HARDWARE AND SOFTWARE CONFIGURATION.

The Instinct Coder 355 is built on the Supermicro AS-A126GS-TNMR server. The blog post lists two AMD EPYC 9575F processors with 64 cores each at 3.3 GHz, eight AMD Instinct MI355X GPUs in the UBB8 form factor with 288 GB of HBM3e memory each, 3 TB of DDR5-6400 ECC system memory, two 960 GB NVMe boot drives, eight 7.68 TB PCIe Gen5 data SSDs, two AMD Pensando Pollara 400 network cards running at 400GbE, and six 5,250 W titanium-level power supplies in a redundant 3+3 arrangement.

The eight GPUs give a single node 2,304 GB of HBM3e. According to AMD's MI355X product page, each GPU is a CDNA 4 part rated at a peak memory bandwidth of 8 TB/s, 5 petaflops of dense FP8 compute and 10.1 petaflops for MXFP4 and MXFP6 matrix operations, with a typical board power of 1,400 W. Those FP4 and FP6 formats are relevant for a node that must serve a large model to dozens of developers at once, because lower-precision inference reduces both memory use and the cost of each generated token, although AMD has not said which precision the GLM-5.3 deployment uses.

On the software side, the platform runs Spectro Cloud's PaletteAI Inference Launchpad, which AMD describes as handling the full AI lifecycle, enterprise governance, local-first model routing and visibility into usage and cost. Support is split between the partners: Spectro Cloud provides full-stack software support, while Supermicro provides three years of next-business-day on-site hardware support.

........

Component	AMD Instinct Coder 355
Server	Supermicro AS-A126GS-TNMR
CPUs	2x AMD EPYC 9575F, 64 cores, 3.3 GHz
GPUs	8x AMD Instinct MI355X, 288 GB HBM3e each
System memory	3 TB DDR5-6400 ECC (24x 128 GB)
Storage	2x 960 GB NVMe boot, 8x 7.68 TB PCIe Gen5 SSD
Networking	2x AMD Pensando Pollara 400, 400GbE
Power	6x 5,250 W titanium PSUs, 3+3 redundant
Local model	GLM-5.3
Optional frontier models	Anthropic Claude, OpenAI GPT, Google Gemini
Developer tools	Claude Code, Cursor, Visual Studio Code
Rated users per node	Up to 85 developers, 50 concurrent

........

··········

INSTINCT CODER 355 VS INSTINCT CODER 325: WHAT AMD CHANGED.

AMD, Supermicro and Spectro Cloud first announced the Instinct Coder concept on August 5, 2026, in a joint press release that described a hybrid, local-first approach to enterprise coding assistants without naming a model. The August solution brief then specified the first configuration on the Supermicro AS-8126GS-TNMR server with eight MI325X GPUs and an AMD Inference Microservices build of GLM-5.2. AMD's current Instinct Coder product page still describes that MI325X configuration and now names GLM-5.3 as its local model.

The 355 keeps most of the host platform unchanged: the same EPYC 9575F processors, the same 3 TB of system memory, the same storage layout and the same power supplies. The changes are concentrated in the accelerators, the server model and the network cards. The MI325X, a CDNA 3 GPU, carries 256 GB of HBM3E with 6 TB/s of peak bandwidth and 2.61 petaflops of dense FP8, according to AMD's product page; the MI355X raises those figures to 288 GB, 8 TB/s and 5 petaflops. The network configuration drops from ten Pensando Pollara 400 cards to two.

........

Specification	Instinct Coder 325	Instinct Coder 355
Server	Supermicro AS-8126GS-TNMR	Supermicro AS-A126GS-TNMR
GPUs	8x MI325X (CDNA 3)	8x MI355X (CDNA 4)
HBM per GPU	256 GB HBM3E	288 GB HBM3e
HBM per node	2,048 GB	2,304 GB
Peak bandwidth per GPU	6 TB/s	8 TB/s
Dense FP8 per GPU	2.61 PFLOPs	5 PFLOPs
Pensando Pollara 400 NICs	10	2
Developers per node	Up to 50	Up to 85
Concurrent users	Up to 30	Up to 50
Cost claim	Up to 70% token savings, about 6-month payback	Up to 75% lower TCO, about 5-month payback
Benchmark claim	Up to 95% of frontier performance	Up to 98% of frontier performance

........

The capacity figures translate to about 10.6 developers per GPU on the 355, against 6.25 on the 325. AMD has not published the throughput measurements, context lengths or latency targets behind the 85-developer rating, so the number describes the vendor's sizing guidance rather than a tested workload. The cost and benchmark claims for the 325 came from Spectro Cloud, and AMD's own product page states that it has not independently verified them and that results may not be typical.

··········

GLM-5.3 AS THE LOCAL MODEL AND THE ROUTING POLICY TO CLAUDE, GPT AND GEMINI.

The local model is GLM-5.3, the open-weight coding model from Z.ai that the platform serves through AMD's inference stack. Data Studios has already covered the model in a different context, in its report on the Cloudflare Workers AI bug that exposed GLM-5.3 responses across customers, which concerned a hosted deployment rather than an on-premises one.

The routing layer decides where each request is processed. In AMD's description, PaletteAI sends requests to the local model by default and allows approved frontier services only when a task justifies them and IT policy permits it. Platform teams define which workloads may leave the on-premises environment, and AMD states plainly that requests routed externally send their content to the selected external service. A company that buys the system to keep source code inside its own walls therefore still has to write and audit the policy that governs those exceptions, since any request routed to Claude, GPT or Gemini leaves the building with its prompt and code context.

The August press release gave more detail on the criteria the routing software can use: enterprise policy, workload complexity, model capability, infrastructure availability and cost. It also described metering and governance features intended to show which teams consume local and external capacity, which is the information a platform team needs to verify whether the promised savings appear in practice.

··········

AMD'S 75% TCO CLAIM FOR INSTINCT CODER 355: THE ASSUMPTIONS AND THE ARITHMETIC.

The headline saving comes from a modeled deployment described in the blog's first footnote. AMD assumes one server with eight MI355X GPUs serving 85 developers who would otherwise each spend $2,000 per month on AI coding and API services, which amounts to about $2.04 million per year. The server is priced at a list price of $590,000 including software and licensing, amortized over three years. Frontier cloud API use continues at 10% of the baseline spend, about $204,000 per year, and data center operating costs are set at $50,000 per year.

Under those assumptions AMD puts modeled annual costs at about $513,500, which corresponds to savings of 74.8% against the $2.04 million baseline, rounded to "up to 75%". AMD gives a payback period of about five months and labels the results as illustrative only.

Data Studios recomputed the model from the listed inputs. One third of the $590,000 server price is about $196,700 per year; adding $204,000 of frontier API spend and $50,000 of operating costs gives about $450,700 per year, roughly $62,800 less than the $513,500 total in the footnote. The blog does not itemize the difference, so the model contains at least one cost that AMD has not described. With AMD's own total, the platform costs about $503 per developer per month against the assumed $2,000, and the three-year cost is about $1.54 million against $6.12 million for the cloud baseline.

The result depends almost entirely on the $2,000 monthly spend per developer. Holding every other AMD input constant, Data Studios calculated how the savings change if a team spends less on hosted AI coding services today. These figures use the listed inputs, which sum to the lower $450,700 total at the original spend level, so they are slightly more favorable than AMD's own model would be.

........

Monthly AI spend per developer	Annual cloud baseline (85 developers)	Modeled annual on-premises cost	Modeled saving
$2,000 (AMD assumption)	$2,040,000	$450,667	77.9%
$1,000	$1,020,000	$348,667	65.8%
$500	$510,000	$297,667	41.6%
$250	$255,000	$272,167	-6.7%

........

On the listed inputs, the break-even point falls at about $269 per developer per month: below that level of current spending, a fully loaded node with a 10% frontier fallback costs more than the cloud services it replaces. The model also assumes that all 85 seats are used, that 10% of the original spend is enough for the requests that still need a frontier model, and that the hardware stays in service for three years. A team of 40 developers on the same node would carry the same fixed costs against less than half the baseline spend.

The 325 generation used a comparable model in its August brief: 50 developers with about $1.2 million of annual cloud inference spend, again $2,000 per developer per month, moving everyday work on premises with a 10% hosted fallback for an estimated saving of about 70%, or roughly $830,000 a year, and a payback of about six months.

··········

GLM-5.3 VS CLAUDE OPUS 5 ON SWE-BENCH VERIFIED: THE 98% CLAIM.

The second claim, "up to 98% of frontier benchmark performance", is based on the Vals AI SWE-bench Verified leaderboard updated on September 1, 2026, with the Mini-SWE-agent scaffold. On that leaderboard GLM-5.3 resolved 95.4% of tasks and Claude Opus 5 resolved 97.0%, which gives GLM-5.3 a relative score of 98.4%. AMD's footnote states that the result applies only to that benchmark and configuration.

The comparison is narrower than the headline suggests. SWE-bench Verified measures whether a model can resolve a curated set of GitHub issues in Python repositories inside a fixed agent scaffold; it does not measure performance on a company's own codebase, on long multi-repository refactors or on languages outside the benchmark set. The reference model is also Claude Opus 5, which is no longer Anthropic's newest model: Anthropic has since released Claude Opus 5.5 and, on November 9, Claude Fable 5.2. The 1.6-point gap therefore describes one leaderboard snapshot rather than the distance between GLM-5.3 and the frontier models a developer can select today.

The figure was also measured on the leaderboard's own infrastructure. AMD has not published SWE-bench results for GLM-5.3 running on the Instinct Coder 355 itself, nor the quantization used on the node, so buyers cannot yet confirm that the local deployment reproduces the leaderboard score.

··········

WHO THE AMD INSTINCT CODER 355 IS FOR AND HOW TO BUY IT.

AMD positions the platform for enterprise software organizations that already spend heavily on hosted coding assistants and APIs and want control over where code is processed, which models are used and what that usage costs, the three points AMD's Dan McNamara listed when the Instinct Coder line was first announced in August. The supported clients are the tools developers already use: Claude Code, Cursor and Visual Studio Code are named in the November announcement, and the August brief also listed OpenAI Codex and Helix.

For teams comparing the system with per-seat subscriptions, the relevant reference point is what developers pay for hosted plans and API usage today; our guide to Claude plans and usage limits sets out the current consumer and business tiers. The AMD model assumes spending well above the price of a single flagship subscription, which is closer to heavy agentic usage billed through APIs.

AMD has not published a purchase price beyond the $590,000 list figure used in its cost model, a general availability date, or regional availability. Interested buyers are directed to AMD, Supermicro or Spectro Cloud sales representatives, and the support terms are fixed at three years of on-site hardware support plus full-stack software support from Spectro Cloud.

··········

AMD INSTINCT CODER 355: CONFIRMED SPECIFICATIONS AND UNVERIFIED CLAIMS.

The hardware configuration is documented in detail: eight MI355X GPUs with 2,304 GB of HBM3e in total, two 64-core EPYC 9575F processors, 3 TB of system memory, a Supermicro AS-A126GS-TNMR chassis and Spectro Cloud's routing software, with GLM-5.3 as the default model and Claude, GPT and Gemini as policy-gated fallbacks. The rated capacity of 85 developers per node is 70% higher than the 50 developers AMD lists for the MI325X-based Instinct Coder 325.

The economic case is less settled. The 75% saving comes from a model whose listed inputs add up to about $450,700 rather than the stated $513,500, whose result depends on each developer currently spending $2,000 a month, and whose break-even falls near $269 per developer per month on those same inputs. The 98% benchmark figure compares GLM-5.3 with Claude Opus 5 on a single SWE-bench Verified snapshot and was not measured on the Instinct Coder hardware. Price, availability dates, the throughput tests behind the 85-developer rating and the precision used to serve GLM-5.3 remain undisclosed, and those are the figures a buyer will need before comparing a node against a year of hosted coding subscriptions.

··········

·····

DATA STUDIOS

·····

datastudios.org