ByteDance's reported model scale, AMD's Taalas acquisition and SAP's squeeze on non-AI spending show how larger AI ambitions are claiming room in hardware plans and corporate budgets.
Artificial Intelligence··Midday
The scale is large, the detail is thin
ByteDance's next artificial intelligence model is described in a Reuters report, attributed to the Financial Times, as approaching the scale of Anthropic's most advanced Mythos system. The copy published by The Star relies on accounts from people described as familiar with the project, and includes no comment from ByteDance or Anthropic. Reuters also says it could not independently verify the report. The concrete development is therefore limited to a reported claim about model scale. No parameter count or training timetable is provided. The report places ByteDance's target beside Anthropic's most advanced system as the point of comparison. It offers no basis for conclusions about the model's capabilities, uses or computing arrangement. Readers have a clearly attributed report that a large-scale model project is under way, rather than a detailed product announcement. The scale ambition is visible; the report leaves the route from project to finished model unspecified.[1]
The model moves into silicon
AMD's agreement to acquire Toronto-based Taalas approaches the computing problem from another direction. Taalas develops model-specific integrated circuits that cast model weights into transistors, instead of repeatedly moving them between memory and a processor. Its first test chip, HC1, was made on TSMC's 6-nanometre process and ran Meta's Llama 3.1 8B model at close to 17,000 tokens per second. Taalas said this was 73 times the rate of Nvidia's H200 while using one-tenth the power; SiliconANGLE notes that the vendor comparison lacked a common workload definition. HC2 targets models of about 20 billion parameters. A finished Taalas part runs only the model for which it was designed. The company says just two of more than 100 layers change for a new design, and its in-house tools can take a design to tape-out in roughly two months. AMD plans to position the chips alongside Instinct accelerators, Helios racks and ROCm software. The undisclosed deal remains subject to customary closing conditions and regulatory approvals. It adds a narrower route built for particular models beside AMD's general-purpose accelerators.[2]
Corporate budgets make room for AI
SAP's decision shows how AI expense can enter a company's everyday allocation of resources. According to an internal email obtained by 404 Media, SAP suspended most travel and hiring last month. Only AI-related travel and hires are exempt from the freeze. The report by Joseph Cox says the company described a need for spending discipline; the decision was taken last month and the email was published on August 6. The development concerns a large software company limiting other activity while making room for AI spending. Together, the three reports reveal different layers of the same resource choice. For ByteDance, the issue is reported model scale; for AMD, adding a company that casts particular models into silicon; for SAP, constraining travel and hiring outside AI. The reports cover different institutions and provide no common price tag. Their connection is that larger AI ambitions reach beyond a model name into chip design, acquisition and operating-budget decisions. Compute expense becomes a visible allocation question that helps determine which tools, purchases and corporate activities receive priority.[3], [1], [2]
