In this briefing
- AMD published a vendor test in which Instinct MI455X ran DeepSeek-R1 under an MLPerf Inference Offline configuration, and AMD added MI450X with 288 GB HBM4 memory to Helios. The test measures offline throughput and cannot replace acceptance tests for production response time, concurrency and power.
- Axelera AI disclosed the Titania architecture and design targets for the AX8 air-cooled reference server. The company plans to put Titania into production in the third quarter of 2027, so the current specifications remain vendor design values and expectations rather than independent results for a purchasable product.
- NVIDIA's filing confirms that Jensen Huang has resigned as President, Chief Executive Officer and a director, and that Colette Kress has become interim Chief Executive Officer. NVIDIA's Board expects operations, customer commitments and the product roadmap to continue without interruption, but that statement is a company expectation rather than a verified delivery result.
AMD brings memory choice within one rack architecture into view
On 12 November, AMD updated the Instinct MI400 Series. AMD says that MI455X reached 21,333 tokens per second per graphics processing unit when it ran DeepSeek-R1 on the AMD Helios rack-scale system under an MLPerf Inference Offline configuration. AMD compared that result with the published NVIDIA Vera Rubin NVL72 MLPerf Inference 6.1 offline result, and AMD reported up to 30% more throughput. A footnote states that AMD Performance Labs completed the test in October 2026 (month 10), that system manufacturers may use different configurations, and that the model, configuration and software version can also change the result.
AMD also introduced Instinct MI450X. The accelerator provides 288 GB HBM4 memory per graphics processing unit, and AMD says that MI450X and MI455X provide the same core compute, memory bandwidth, scale-up and scale-out connectivity, and ROCm software foundation. AMD associates the additional memory headroom in MI455X with larger models, longer contexts or higher concurrency. The announcement does not fully disclose MI450X pricing, single-accelerator power, a definite delivery date or an enterprise service level.
For enterprise adoption, a memory choice within the same rack architecture can reduce some software and integration differences, but an offline-throughput result cannot answer time to first token, sustained concurrency, network bottlenecks, power or total cost. Only a test using the same model, quantisation method, context length, batch, software version and power boundary can turn the capacity difference between MI450X and MI455X into a comparable deployment result.
Axelera AI makes existing air-cooled data centres a constraint for a future architecture
On 12 November, Axelera AI disclosed its third-generation Titania chiplet architecture. The company says that Titania will be offered as standard PCIe cards and as boards that conform to the Open Compute Project specification. The planned Voyager toolchain will support tensor, expert and pipeline parallelism, as well as speculative decoding and key-value-cache offload and disaggregation. The company also states that Titania is planned to enter production in the third quarter of 2027 and that customer deployment commitments extend into early 2028. Titania therefore remains an architecture and reference design; it is not a product that is already in mass production or on sale.
Axelera AI describes the AX8 reference server as a 4U air-cooled server containing eight system-in-package accelerators. Vendor design values include 20 PFLOPS of 8-bit floating-point compute, 1 TB of accelerator memory, 19.6 TB per second of memory bandwidth and a typical power draw between 5 and 6 kilowatts. The company also says that each package uses 128 GB LPDDR6 memory and that one server can hold a 4-bit floating-point model with one point six trillion parameters, expressed as 16,000 groups of one hundred million parameters, while retaining headroom for the key-value cache. The public material does not provide independent Titania silicon testing, pricing, time to first token, concurrency conditions, offline security certification or the term of enterprise support.
For enterprise adoption, the 4U form factor and design power of between 5 and 6 kilowatts for AX8 make reuse of an existing air-cooled data centre a specific target for further verification, but an architecture planned to enter production in the third quarter of 2027 cannot be treated as a current procurement option. The reference design provides one future example of lower power, a smaller form factor and local model capacity, but one vendor target does not establish a continuous trend towards smaller enterprise inference infrastructure.
NVIDIA's leadership transition adds supplier continuity to procurement comparisons
On 12 November, NVIDIA filed a Form 8-K with the U.S. Securities and Exchange Commission. The filing confirms that the NVIDIA Board accepted Jensen Huang's resignation as President, Chief Executive Officer and a director on 11 November, with immediate effect. The Board appointed Executive Vice President and Chief Financial Officer Colette M. Kress as interim Chief Executive Officer, and Kress continues to serve as Chief Financial Officer during the interim appointment. The Board also appointed Harvey C. Jones as Chair and has begun a search for a new Chief Executive Officer.
NVIDIA's announcement on the same day says that the Board expects the company's operations, customer commitments and product roadmap to continue without interruption. The statement reflects the Board's current expectation, while the filing does not explain the reason for the resignation and does not report a product delay, a change in supply or a change in customer commitments. The public material can therefore confirm the leadership transition and the formal search, but it cannot support a further conclusion about specific hardware-delivery results.
For enterprise adoption, a leadership transition makes supplier governance, roadmap responsibility and support continuity separate procurement dimensions, but the transition cannot by itself prove that delivery has been interrupted. An enterprise assessment of supplier continuity needs to distinguish governance facts confirmed by a regulatory filing, the company's expectation for future operations, and delivery commitments that can be verified in contracts, supply plans, software support and partner responsibilities.
What to watch next
- Can AMD disclose MI450X single-accelerator power, pricing, a definite delivery date and enterprise-support conditions, and can independent tests reproduce MI455X performance under interactive and concurrent workloads?
- Can Axelera AI deliver Titania silicon as planned in the third quarter of 2027, and can production systems provide independent results for performance, pricing, offline security, software compatibility and after-sales support?
- When will NVIDIA appoint a permanent Chief Executive Officer, and can later NVIDIA disclosures continue to confirm the execution status of the product roadmap, supply arrangements and customer commitments?
Sources and verification
Golden Data has edited this briefing from the public materials listed above. The original sources govern facts and figures. The enterprise relevance sections are Golden Data editorial analysis and do not constitute an endorsement of any third-party product.
← Back to AI Daily Briefing