Saturday, 15 August 2026 · Daily Market Intel

Market IntelSat, 15 Aug

The model stopped being the constraint; power and geography took over.

Physical limits focus · post-model leverage · siting

Tech Breakthroughs
- Better agents without a bigger base: Z.ai's GLM-5.3 lifts Terminal-Bench from 4.6 to 28.3 — MarkTechPost
  Z.ai shipped GLM-5.3 on the same 743B base model as GLM-5.2. Every gain came from scaled post-training across more task environments; DeepSWE rose from 46.2 to 66.9, and the weights follow in about two weeks.
  Positive-sum angle: Gains that come from environments rather than parameters travel cheaply. Teams building agent harnesses, evaluation suites, and task simulators all benefit from the finding that the training environment is the scarce input, and open weights two weeks later spread the result.
  What's the impact: SEA teams without frontier compute should invest in task environments and evaluation harnesses for their own domains: Bahasa customer service, cross-border logistics, SGX filings. The environment is the defensible asset now; the base model arrives free a fortnight later.
Emerging Industries
- Batam has the latency but not the water: 60% of Indonesia's power still burns coal — TNGlobal
  BMI reports that private-capital platforms including Digital Edge, PDG, and DCI hold over 60% of Indonesia's data-centre pipeline. Batam offers sub-5-millisecond latency to Singapore and draws its water from rainfall alone.
  Positive-sum angle: A constraint map creates markets. Every megawatt, cubic metre, and fibre route Batam lacks prices a business for someone else: Johor operators, water-reuse vendors, subsea route owners, and the local co-location firms who top out at two megawatts today.
  What's the impact: Indonesian and Singaporean builders should sell into the gap rather than the boom. Cooling-water reuse, grid-interconnect engineering, and Batam fibre are procurement lines being written now, while co-location capacity waits years in the power queue.
- AI lifts the fabs and squeezes everyone else: Malaysia's chip exports near $110B — TNGlobal
  HSBC finds Malaysia's AI-exposed chipmakers and data-centre suppliers pulling ahead while consumer-electronics firms absorb higher memory prices. Semiconductor exports reached nearly $110 billion in 2025, 23% of GDP, on 13% of global assembly, testing, and packaging.
  Positive-sum angle: A widening gap inside one sector opens room beneath it. Memory-efficient design shops, test-equipment vendors, and engineering academies all gain as Malaysia works toward its 60,000-engineer target, and the talent that trains there serves Penang, Singapore, and Ho Chi Minh alike.
  What's the impact: Malaysian and Singaporean operators should hedge the helium and memory lines in 2026 contracts, not 2027. The National Semiconductor Strategy funds the training, so wage parity against Singapore decides who keeps the engineers it produces.
Open Source Projects
- Tool-calling now fits in 28MB of RAM: Cactus Compute's Needle 2 ships as a 14MB binary — MarkTechPost
  The open 45M-parameter model calls tools on a Raspberry Pi 5 at 500 tokens per second and on sub-$200 Android phones at 300 to 700, leading Seal-Tools out-of-domain at 28.7%, with prebuilt binaries across six platforms.
  Positive-sum angle: Intelligence small enough for a $200 phone widens who can build. Voice-to-action appliances, offline field apps, and receipt extraction stop carrying a cloud bill, so developers serving low-ARPU users compound alongside the ones serving Silicon Valley.
  What's the impact: Founders across Indonesia, Vietnam, and the Philippines should design offline-first for the sub-$200 Android base rather than treating it as a degraded tier. Local routing that escalates to the cloud only on low confidence cuts latency and the data-residency argument at once.
Business Models
- Capital in one country, engineering in another: US startups took 64% of 2025 global funding — TNGlobal
  Beyond Border, which advised roughly 6,500 founders in twelve months with about half from Asia, describes a structure now common: founders and commercial teams in the US, product and engineering in Bangalore, Gurugram, or Singapore.
  Positive-sum angle: Splitting the company keeps both ends. US capital and enterprise relationships reach founders whose engineering salaries, training, and equity stay in Asian cities, so the talent pool Singapore and Bangalore built keeps compounding at home while the revenue arrives from abroad.
  What's the impact: SEA founders should set the immigration and entity timeline at incorporation, not at Series A. Visa scrutiny runs long enough that a US commercial presence takes the better part of a year to stand up, and the round closes faster than that.

Never miss a SEA deal.