Monday, 24 August 2026 · Daily Market Intel

Market Intel Mon, 24 Aug

Intelligence goes free; memory and legibility collect the toll.

Free weights focus · scarce memory · agent legibility

Tech Breakthroughs
- The base is free, the moat is not: Harvey post-trains Kimi K3 for law — MarkTechPost
  
  Harvey and Fireworks AI ran asynchronous reinforcement learning on Moonshot's open Kimi K3 across roughly 150 NVIDIA B300 GPUs for two months. Tenet nearly doubles completed tasks on Harvey's Legal Agent Benchmark. No weights, no API.
  
  Positive-sum angle: Open Chinese weights become shared industrial stock. Every vertical that post-trains on K3 proves a recipe the next one copies, and Fireworks sells the same pipeline to the next buyer, so the cost of a domain-grade agent falls each time someone pays it.
  
  What's the impact: SEA vertical-software founders should stop pricing a foundation model into the plan and price the post-training run instead: two months of rented B300s and a proprietary evaluation set. The defensible asset in Jakarta or Manila is the benchmark, not the base.
- China's chips handle cheap inference; coding still needs Nvidia, and tokens hit 140 trillion a day — SCMP
  
  Chinese inference firms run low-value workloads on domestic accelerators and reserve scarce Nvidia chips for coding and agent tasks, where quality pays. Daily token calls passed 140 trillion in March, up more than a thousandfold since early 2024.
  
  Positive-sum angle: The split creates a two-tier market with room in both halves. Domestic silicon absorbs the bulk-inference floor and funds its own iteration, while scarcity at the top pulls capital into serving-layer efficiency work that every buyer downstream inherits.
  
  What's the impact: SEA operators buying inference should price the two tiers separately and route bulk summarisation, transcription and classification to the cheap floor now. Reserve frontier capacity for coding agents, and lock Johor and Batam contracts before the coding tier reprices.
Emerging Industries
- Phones earned $67.8B and returned almost nothing: Samsung's chip division took 97% of the profit — TNGlobal
  
  Samsung posted $205.6 billion in H1 revenue and $80 billion in net income. Semiconductors delivered $96.2 billion of operating profit, about 97 percent of the group total, while consumer electronics turned $67.8 billion of revenue into $1.4 billion.
  
  Positive-sum angle: Memory scarcity pays the region that builds it. Samsung's Vietnamese subsidiaries booked $35.2 billion, 17.1 percent of group revenue, and the new Thai Nguyen semiconductor plant moves Vietnam up from assembly, training engineers who will staff the next supplier.
  
  What's the impact: SEA hardware and cloud founders should budget for DRAM and HBM priced by AI demand, not by handset cycles, through 2027. Vietnamese industrial-services and precision-component teams should sell into Thai Nguyen now, while the supplier list is still being written.
Open Source Projects
- A 753B model on one desktop GPU: FreeToken trades VRAM for 512GB of host memory — MarkTechPost
  
  UC Berkeley and UT Austin researchers released FreeToken under Apache-2.0, a mixture-of-experts serving engine that splits work between GPU and CPU. It runs 753B GLM-5.2 on one workstation GPU and 35B at 39.3 tokens per second on an 8GB laptop.
  
  Positive-sum angle: Cheap DRAM substitutes for scarce VRAM, which moves large-model access to anyone who can buy memory. University labs, ministries and small teams across the region gain a local deployment path, and every optimisation they contribute back lowers the floor for the next workstation.
  
  What's the impact: SEA teams holding data they cannot send offshore, in health, defence procurement and banking, should test FreeToken on a 512GB workstation before signing another private-cloud contract. The compliance case for on-premise inference now carries a hardware bill instead of a lab budget.
Business Models
- Agents will grade your storefront before customers do: Vercel's free audit runs 118 checks — MarkTechPost
  
  Vercel launched Is Agentic, a free tool scoring how well AI agents discover, access, understand and pay on a public website. The methodology comes from Ora and weights usability at 40 points, access 30, discovery 20 and payments 10.
  
  Positive-sum angle: A public, free scoring standard turns agent-readiness into a checklist instead of a consulting engagement. Small merchants get the same audit as large platforms, and each site that fixes its structured data makes the agent layer more reliable for every other seller on it.
  
  What's the impact: SEA e-commerce and travel operators should run the audit on their storefronts this quarter, while a passing grade is still rare enough to matter. Agentic checkout is arriving on Alipay and Mastercard rails, and the site has to be readable before the payment works.

Never miss a SEA deal.