In this briefing
- Z.ai released GLM-5.3-Flash with 320 billion total parameters and 18 billion active parameters, native multimodality and a 1 million-token context, and made its weights available under the MIT licence; official API pricing is currently discounted until 9 September.
- vLLM 0.28.0 brings weight offloading, E/P/D disaggregation and tiered KV cache into a formal release, with CUDA, ROCm, CPU and XPU artefacts. The same version also contains incompatible changes such as dependency upgrades and removed interfaces.
- AWS and NVIDIA plan to deploy an additional 2 million Blackwell Ultra, Rubin and Rubin Ultra GPU accelerators in 2027–2028. This is a future-capacity plan, not resources that are already online, and no transaction value or cloud pricing has been disclosed.
Model layer: designing inference demand into the architecture
Z.ai released GLM-5.3-Flash on 26 August. Official material describes it as a natively multimodal model with 320 billion total parameters and 18 billion active parameters, using a hybrid of linear and sparse attention and supporting a 1 million-token context, image and video input, file handling, tool calls, JSON output, streaming responses and caching. The model weights are published under the MIT licence through Z.ai’s official Hugging Face account; the model card lists Transformers, vLLM, SGLang, TokenSpeed and KTransformers as runtime routes.
The pricing page lists temporary rates per 1 million tokens of $0.075 for input, $0.015 for cached input and $0.25 for output, with the discount ending at 24:00 (UTC+8) on 9 September 2026. Official material also gives reduction figures for attention computation and KV cache relative to GLM-5.3, but these are vendor estimates. Public information does not provide minimum memory, quantisation conditions, time to first token, throughput, concurrency, power or an independent production benchmark in the same disclosure, and the model’s thinking mode cannot currently be disabled.
For enterprise implementation, fewer active parameters and hybrid attention provide another cost lever, but 18 billion active parameters cannot be converted directly into single-machine deployability or serving capacity. Actual resource needs still depend on precision, quantisation, context length, batch size, output length and quality constraints. The MIT licence and existing runtime frameworks broaden private-deployment options, but they do not amount to verified hardware adaptation, enterprise support or security and compliance certification. Temporary API pricing likewise cannot be used directly as a long-term total cost of ownership.
Runtime layer: KV cache begins to span storage tiers
The vLLM project released version v0.28.0 on 26 August. Model Runner V2 adds E/P/D disaggregation, weight offloading and multi-layer MTP KV cache support. Tiered KV cache adds disk offloading, secondary-tier managers implemented through external modules, partial-load results, tiering metrics and a standard CPU layout independent of the parallelism method.
The release provides images or installation artefacts for CUDA 13, CUDA 12.9, ROCm, CPU and XPU, extending deployment entry points across NVIDIA, AMD, Intel and general-purpose CPUs. The notes also list breaking changes: bitsandbytes support moved to an external plugin, the Transformers dependency was upgraded, and calculate_kv_scales and override_attention_dtype were removed. The Kimi-K3 acceleration and memory-saving figures on the page come from project-contributor tests under specific model and configuration conditions; they are not independent production benchmarks and cannot be generalised to other models, hardware, contexts or concurrency conditions.
For enterprise implementation, cache becomes a service-layer asset that can be organised across GPU, CPU and disk, while tiering and disaggregated execution extend capacity optimisation beyond adding accelerators to include memory hierarchy and service topology. That may improve the use of scarce accelerators, with added variables in data movement, tail latency, failure domains, monitoring and version compatibility. An offloading capability in the runtime does not mean that a particular workload will necessarily cost less; model-access patterns, storage bandwidth, network topology and service-level objectives still determine the outcome.
Supply layer: 2 million GPUs remain a future plan
On 26 August, AWS and NVIDIA announced a plan to deploy an additional 2 million NVIDIA GPU accelerators across AWS’s global infrastructure in 2027–2028, covering Blackwell Ultra, Rubin and Rubin Ultra. The scope also includes Vera CPU, NVLink Fusion and NVHBM, plus integration with AWS Nitro and Elastic Fabric Adapter. The companies also propose AI infrastructure for the United States government containing 100,000 GPUs and meeting IL6 or higher security requirements.
The announcement describes an expansion route for the next two years; it does not mean that 2 million GPUs have been purchased, delivered or made available to customers. The official release does not disclose the transaction value, regional or timetable breakdown, specific instance types, availability or prices. TechCrunch’s contemporaneous report also describes planned deployment.
For enterprise implementation, the plan for 2 million GPUs strengthens the roadmap signal that hyperscale cloud supply may continue expanding, but it is not enough to infer lower GPU cloud-service prices or to replace evidence about supply for a particular region and instance. Procurement conditions still depend on actual region, instance shape, network and storage support, delivery cadence, service levels and pricing. This is not a substitute for model and runtime efficiency: a larger infrastructure pool can carry more workloads, while the resources used by each inference call still depend on model architecture and serving engineering.
Integration layer: same-day patches expose version boundaries
Hugging Face released Transformers v5.16.0 and v5.16.1 in sequence on 26 August. Transformers v5.16.0 replaced the legacy tensor-parallel backend with an implementation based on DTensor; about two hours later, v5.16.1 restored the previous compatibility behaviour, established a deprecation period and added support for GLM-5.3-Flash. The model release and framework adaptation therefore connected on the same day.
The GLM-5.3-Flash model card lists several local runtime frameworks, and vLLM 0.28.0 provides cross-hardware artefacts; the public material for the two has not formed a jointly validated enterprise deployment matrix.
For enterprise implementation, an ‘open model’, a ‘supported framework’ and a ‘formal runtime release’ are three different states; the same-day restoration of compatibility also shows that a support list does not remove constraints from dependency versions and migration paths. A change to any layer—model weights, transformer version, inference engine, quantisation format, driver or hardware backend—can alter correctness, performance and operability. Taken together, the day’s three types of update suggest that scaling should be assessed across a version-coordinated service stack. The capabilities form a reproducible deployment result only after the version combination, data conditions and service objectives are fixed; upstream compatibility restoration or incompatible changes also enter validation scope, rollback conditions and lifecycle cost.
What to watch next
- Independent tests of GLM-5.3-Flash time to first token, throughput, concurrency, power and quality under the same precision, context, batch and hardware conditions.
- Stability, tail latency and migration cost for vLLM tiered KV cache and E/P/D disaggregation under mixed hardware, long contexts and sustained concurrency.
- The regions, instance specifications, prices, network and power support behind the AWS and NVIDIA plan, and its actual delivery cadence during 2027–2028.
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