Weekly Brief for IT and Development: Chips, AI Agents, and EV Charging Moves
Big signals this week: South Korea is localizing EV charger converter modules to cut China exposure. ByteDance is reportedly exploring Samsung as a manufacturing partner for an in-house AI chip. Foundries are pulling back from mature nodes, while MCU vendors are hiking prices. Meanwhile, AI agents are starting to pressure traditional SaaS and cloud spend models.
Supply chain reset: EV charger modules go local in Korea
South Korea's push to produce EV charger converter modules domestically points to a clear trend: reduce single-country risk in critical components. If you build charging software, firmware, or field services, expect spec changes and more vendor diversity in 2026 bids.
- Action: keep hardware abstraction layers loose-treat charger comms (OCPP, ISO 15118, PLC stacks) as swappable backends.
- Action: extend CI to validate across multiple AC/DC and DC/DC module vendors, especially for thermal throttling behavior and fault codes.
AI chips: ByteDance looks to Samsung amid Nvidia constraints
Reuters reports that ByteDance is developing an AI chip and has discussed manufacturing with Samsung. The goal is clear: secure inference and training capacity without relying only on US suppliers under export limits. Expect more hyperscalers and big app platforms to try custom silicon to manage cost and availability.
- Action: design your ML stack to be portable-support CUDA and ROCm, keep ONNX export clean, and avoid vendor-locked ops.
- Action: build kernels in Triton or TVM where possible so you can retarget if procurement shifts.
Source: Reuters report on ByteDance-Samsung chip talks
Cloud economics: AI agents hit a commercial moment
Teams are moving from demos to revenue-backed agents. The catch: token burn, retrieval quality, and orchestration costs can erase margins fast. The winners are adding caching, short-context prompts, and tight tool use with strict guardrails.
- Action: measure dollar-per-task, not just latency-gate production rollout based on unit economics.
- Action: standardize evals for hallucination, tool misuse, and PII leaks; run them per model update.
- Action: keep a fallback path (rules or classical code) for high-risk steps.
If your org is upskilling for this shift, see our focused learning tracks: AI courses by job role.
Foundries, OSAT, and packaging: capacity gets reshuffled
Several major foundries are exiting or reducing mature-node capacity as pricing strengthens. Top OSAT rankings for 2025 put five Chinese firms in the top 10, signaling tighter competition in advanced packaging and test. MCU vendors are raising prices on packaging cost inflation, which will hit embedded and IoT BOMs first.
- Action: lock in second-source MCU SKUs now; prioritize pin-compatible options to avoid PCB spins.
- Action: re-quote your packaging-heavy parts (QFN with large pad counts, SiP) and update cost models for 2H26.
- Action: for ASICs, plan earlier DFT and probe strategy; tester queue times will lengthen as overflow work hits mid-tier houses.
Displays and devices: OLED and Mini LED squeeze
Samsung Display is reviewing capacity for foldable OLEDs reportedly tied to big-brand demand. At the same time, pricing pressure on large OLED TVs continues as Chinese Mini LED sets gain traction. For app and TV OS developers, this means stricter HDR targets and more variability in dynamic refresh behavior across panels.
- Action: expand device labs with diverse panel types; test HDR10/HLG tone mapping, ABL behavior, and PWM flicker sensitivity.
- Action: instrument jank and frame pacing; foldables need layout policies that don't spike overdraw on transitions.
Networking and data movement: 102.4 Tbps era
Cisco announced 102.4 Tbps-class switch silicon and new systems aimed at broader AI infrastructure. For platform teams, the takeaway is simple: east-west traffic will keep climbing, and oversubscription targets will get tighter in training pods.
- Action: prepare for 800G and plan cabling now; qualify reliability and thermal headroom before scaling pods.
- Action: move telemetry closer to the NIC; sample fine-grained drops, ECN marks, and queue depth to catch microbursts.
Energy and EV charging: Taiwan builds, policies lag
Tesla is expanding charging coverage in Taiwan, even as policy questions hang in the air. Taiwan's renewable backup requirements are also driving more energy storage, which will push utilities and data centers to revisit uptime strategies and grid contracts.
- Action: model site loads with worst-case charger concurrency; test throttling logic with real breaker curves.
- Action: for DCs, integrate storage telemetry with incident response-treat it like another critical subsystem.
Safety and robotics: trust-first frameworks
Boston Dynamics highlighted a safety-first playbook for humanoids. For teams building physical AI, regulations and public trust move slower than demos, so documentation and fail-safes need to be production-grade from day one.
- Action: map your controls to a recognized framework such as the NIST AI Risk Management Framework.
- Action: maintain a change log that ties software updates to safety tests and incident drills.
What to do this week
- Qualify at least two EV charger module vendors and add automated regression tests for fault reporting.
- Make your ML pipeline hardware-agnostic: clean up ONNX exports, dual-stack CUDA/ROCm, and track kernel hotspots.
- Re-price MCUs and packaging-heavy parts; secure buffer stock where single-sourced.
- Pilot agent workloads with strict unit economics and caching; ship only where cost per task is predictable.
- Plan for 800G optics in 2H26; validate thermal envelopes in your densest racks.
- Expand panel testing coverage for HDR and foldables; monitor flicker and motion artifacts.
The pattern is consistent: diversify suppliers, keep software portable, and measure costs at the task level. Small moves now will save rebuilds later.
Your membership also unlocks:
