Huawei's AI Stack, Mythos Access and Memory Bottlenecks

The Pretrained Pod1h 10m
Pierce Freeman and Richard Diehl Martinez in blue and white tops against black, alongside the blue and white headline "AI'S HARDWARE BOTTLENECKS".
0 comments · 0 votesOpen discussionClose discussion
Sign in to join the discussion

    Pierce Freeman and Richard Diehl Martinez discuss reports about DeepSeek and Huawei's Ascend platform, asking how much of Nvidia's advantage comes from its software ecosystem rather than individual chips. Their central distinction is between manufacturing hardware and making existing workloads run efficiently on it. Compatibility, developer tooling and communication between acceleratorsAI compute infrastructure is the hardware, facilities, networks, power, cooling, storage, and software used to train and run AI models. remain important even when programs can be translated.

    The discussion of restricted Mythos access connects scarce capacity, cybersecurity and government influence. The hosts consider reported requests around access rather than establishing a binding government ban. AISI's controlled cyber evaluationsEvaluation is the systematic assessment of a system's behaviour or results against explicit criteria. demonstrate progress in sustained attacks, but do not establish that every well-defended real network can be compromised.

    Memory demand leads to a discussion of high-bandwidth memory and the state needed for long-context inferenceAI inference economics examines the revenue, cost, utilization, pricing, and margins involved in serving trained models.. The hosts debate whether compression and architectural changes could reduce costs. Their personal hardware-price experiences and forecasts of supplier bargaining power are not a verified market-wide price series, and model context limits should not be confused with permanent conversational memory.

    The Manus acquisition dispute prompts a debate about national jurisdictionGovernance defines decision-making authority, rules and accountability within a system or organization. over AI expertise, corporate location and investment. In the closing section, the hosts offer competing views of global opportunity: paid compute could disadvantage talented developers in lower-income regions, while cheaper prototyping could reduce the capital needed to start a company. These are possible outcomes, not settled predictions.

    Original YouTube thumbnailWatch on YouTube

    Video topic coverage

    • Agents & workflows11.1
    • Coding & tools44.4
    • Models & learning55.6
    • Infrastructure & compute100
    • Research & evaluation38.9
    • Safety & security55.6
    • Business & productivity66.7
    • Creative & media0
    • Robotics & perception0
    • Society & policy83.3

    Share this video

    Pierce Freeman and Richard Diehl Martinez in blue and white tops against black, alongside the blue and white headline "AI'S HARDWARE BOTTLENECKS". With WWW.ARTIFICIAL-INTELLIGENCE.VIDEO, 9 May 2026 and duration 1h 10m.

    AI competition depends on software, interconnects and memory as well as chips. The hosts explore portability to Huawei hardware, restricted cyber-model access and how compute costs and regulation could shape global entrepreneurship.