FoundersFeed

Photonics industry hitting a weird supply chain wall: legacy 90nm and 45nm fab capacity is now the critical bottleneck. These older process nodes are essential for manufacturing photonic integrated circuits (PICs) because they balance cost, yield, and the specific geometries needed for waveguides and optical components.

The issue? Most semiconductor fabs have moved to cutting-edge nodes (5nm, 3nm) or shut down older lines. But photonics doesn't need transistor density—it needs mature, stable processes with good control over feature sizes in the 45-90nm range for efficient light manipulation.

This creates a supply crunch: companies building optical interconnects, LiDAR systems, or quantum photonics are scrambling to secure fab time on equipment that's considered obsolete in the mainstream chip world but is gold for optics. The mismatch between what's profitable for foundries versus what photonics needs is creating real production delays.

Expect to see either: (1) dedicated photonics fabs emerging, (2) existing players like GlobalFoundries or Tower Semiconductor carving out legacy capacity, or (3) vertical integration where photonics companies buy or lease old fab lines outright. The economics are tricky but the demand is real—especially with AI clusters needing faster optical I/O and autonomous vehicles requiring cheap LiDAR.

Quantum computing = misnomer. It's quantum processing, not general-purpose computing. Think of it as a specialized accelerator for specific problem classes (optimization, simulation, cryptography), not a CPU replacement. You wouldn't drive a boat on a highway—same logic applies here. Classical computers handle logic gates and sequential ops. Quantum processors exploit superposition and entanglement for parallel state exploration. Different tools, different jobs. The hype cycle conflates "quantum supremacy" with "replaces your laptop," which is fundamentally wrong. It's a co-processor for narrow domains where exponential speedup matters.

Photonics chips don't need bleeding-edge process nodes like 3nm or 5nm. 90nm and 45nm nodes are totally sufficient for current photonic integrated circuits.

Why? Photonic components (waveguides, modulators, photodetectors) are physically larger than transistors. They operate with light wavelengths around 1550nm, so sub-10nm lithography doesn't help. The critical specs are optical loss, modulation bandwidth, and coupling efficiency—not transistor density.

45nm nodes give you enough precision for passive photonic structures while keeping costs way down. You're not trying to pack billions of transistors here. You're routing light and converting it to electrical signals.

This is why companies like Intel and GlobalFoundries use mature nodes for silicon photonics. The real engineering challenge is hybrid integration with electronics, not shrinking the photonics themselves.

Photonics chip architecture is basically 99% silicon substrate with just a 1% thin film layer on top doing the heavy lifting. The bulk silicon acts as the mechanical foundation and waveguide base, while that microscopic top layer handles the actual light manipulation—phase modulation, optical switching, whatever. This is why photonics scales so well economically: you're leveraging mature silicon fab processes for almost everything, then depositing specialty materials (like silicon nitride or III-V compounds) only where photon-level control is needed. The manufacturing cost curve follows semiconductors, but the physics is pure optics. That 1% is where all the magic happens.

PyTorch now compiles directly to photonic hardware instructions. Instead of running matrix ops on silicon, you're generating control signals for optical computing systems—manipulating light interference patterns for tensor operations. The compiler translates standard torch.nn modules into photonic circuit configurations, handling the mapping between digital weights and analog optical modulators. This bridges the gap between training in software and inference on photonic accelerators, which theoretically hit way lower power consumption for linear algebra at scale. Key challenge: quantization noise and thermal drift in optical components, but if they nail the error correction, this could actually make photonic ML practical beyond lab demos.

Agents don't sleep — and that's breaking traditional compute architectures.

Mohamed Awad (EVP Cloud AI at Arm) nails the core bottleneck: CPUs are now always-on coordinators in agentic systems, not just interactive workload handlers.

The problem? Agent sprawl. One agent spawns 10, those spawn 100, then 1000 — recursive task delegation at scale. Every spawned agent needs orchestration, state management, and scheduling. That's all CPU work.

Energy becomes the real constraint. Every milliwatt burned on CPU overhead is a milliwatt you can't allocate to accelerators (GPUs/TPUs) or serving more agents. It's a zero-sum game at datacenter scale.

Implication: We need ultra-efficient CPUs purpose-built for agent coordination, not general-purpose compute. Think low-power, high-concurrency, optimized for context switching and async task graphs.

Arm's pitch here is obvious — their low-power architecture could dominate this coordination layer while leaving heavy lifting to accelerators. But the real insight is architectural: agentic AI demands a new compute hierarchy where CPUs are lightweight orchestrators, not heavyweight processors.

Mohamed Awad (Arm EVP Cloud AI) pushes back on "CPU is dead" hype:

The real story isn't GPU vs CPU—it's that AI agents are hammering EVERYTHING. CPUs, storage, networking, every layer of the stack is getting crushed under token generation and agent orchestration load.

"The CPU market is growing massively because we need to serve up intelligence."

He calls out the narrow thinking of just comparing CPU-to-GPU ratios in a single blade or rack. AI workloads don't live in isolation—they're distributed, data-hungry, and require coordination across the entire infrastructure.

As agents multiply and distill massive token streams, the bottleneck isn't just compute—it's the whole pipeline. CPUs are the backbone of that coordination layer, and demand is skyrocketing.

Claude 3.7 Sonnet just dropped with extended thinking mode. Architecture runs internal reasoning chains before responding - literally watching it think through problems step-by-step like GPT-o1 but with Anthropic's safety rails.

Key difference: You can see the thinking process in real-time through the API. Costs 3x more tokens since you're paying for both the reasoning and the output, but benchmark improvements are solid for math, coding, and multi-step logic.

Best use case: Complex debugging sessions where you need the model to actually reason through edge cases instead of pattern-matching from training data. The thinking tokens add up fast though - budget accordingly.

China's AI scene is hitting parity with the US, but nobody's paying attention to what's actually shipping from Shenzhen labs.

David Li from Shenzhen Open Innovation Lab points out the obvious: Western devs obsess over Trump tweets while Chinese teams are quietly iterating on models, deploying at scale, and building infrastructure that rivals anything coming out of SF.

The asymmetry is wild - we track every OpenAI blog post but ignore the technical output from Alibaba's DAMO Academy, Tencent's AI Lab, or ByteDance's research teams. Xi's policy statements don't matter as much as the actual code being pushed to production.

If you're only reading English-language AI news, you're missing half the global model development cycle. The real competition isn't happening on Twitter - it's in training runs, dataset curation, and inference optimization that Chinese labs are running 24/7.

China's hardware scene is mass-producing AI talking toys at insane speed. David Li from Shenzhen Open Innovation Lab breaks down the economics: $5 chip + $10 flat-rate LLM token plan = instant product. The entire pipeline from component sourcing to Amazon listing has zero friction.

The barrier to entry is basically nonexistent. Shoot a demo video, slap together an Amazon page, you're live. No R&D moat, no IP protection, pure manufacturing velocity.

This is what happens when you combine Shenzhen's supply chain density with commoditized AI APIs. The race isn't about innovation anymore, it's about who can ship fastest and undercut on price. Wild West for consumer AI hardware.

Free Character & Story Generator tool just dropped

Basically takes your character concept and auto-generates narrative context around it. Useful if you've got character design locked but haven't figured out the plot/backstory yet.

Pretty straightforward workflow - feed it character details, get story scaffolding back. Could be handy for game devs, writers, or anyone doing narrative design who needs to bootstrap story elements quickly.

Worth checking out if you're stuck at the "cool character, now what" phase of a project

Arm's Cloud AI EVP highlights a critical supply chain reality: AI infrastructure isn't just bottlenecked by GPUs anymore. The real blockers are mundane components like capacitors and passive electronics.

A hardware team couldn't ship their AI board because they couldn't source basic capacitors. These aren't exotic parts—they're commodity components you'd assume are always available. But demand for AI hardware has created unexpected shortages in the entire supply chain stack.

The takeaway for builders: when planning AI infrastructure deployments, factor in lead times for passive components, PCB assembly capacity, and skilled technicians who can actually build and debug these systems. The chip shortage narrative misses half the problem—it's the entire BOM and labor pipeline that's constrained.

This is why hyperscalers with deep supply chain relationships and vertical integration have such a massive advantage. Startups trying to build custom AI hardware are getting blocked by $0.05 capacitors, not $10k GPUs.

We're living through a massive regression in software UX quality. The industry is racing to ship features faster, not better.

Case in point: GitHub Actions still displaying bugs in production. This isn't just about one broken UI element - it's a symptom of the velocity-over-quality mindset that's become normalized in tech.

The real cost? Technical debt compounds faster than teams can pay it down. Users get trained to expect broken experiences. And the bar for "shippable" keeps dropping.

The irony: tools built for developers (like GitHub) are suffering from the same rushed release cycles they enable. When the infrastructure layer starts showing cracks, it's a canary in the coal mine for the entire ecosystem.

Cosmos EVM is getting exploited right now.

Two brutal truths exposed:

1. Most DeFi isn't actually decentralized - when validators can emergency-halt chains mid-hack, you're just trusting a permissioned cartel with extra steps.

2. Turing-complete smart contract platforms = infinite attack surface. Every DeFi protocol is a nested stack of composable vulnerabilities waiting for the right exploit path.

The complexity tax is real: bridges, oracles, AMMs, yield aggregators - each layer multiplies risk exponentially. One logic bug in a dependency chain can cascade into total fund loss.

$BTC's design philosophy wins here: fixed script opcodes, no Turing completeness, predictable state transitions. Less programmability = less exploitability.

The irony: chains marketed as "censorship-resistant" routinely pause when real money is at stake. That's not a bug, it's revealing the actual trust model.

US Treasury just dropped a notice that could hit crypto infrastructure hard. They're threatening secondary sanctions on "any entity" facilitating money laundering or sanctions evasion for Iran.

The scary part for devs: no hard limit excluding noncustodial protocols or developers. The language is intentionally vague - "any source" and "every economic lifeline" means they're leaving the door open to go after DeFi protocols, mixers, or even core contributors if they decide your code is helping Iranian regime finance.

Most of the focus is on traditional intermediaries - Iranian exchanges, front companies, payment processors. But the lack of explicit exemption for decentralized infrastructure is the real signal here. This is Treasury saying "we might come for your protocol too if we feel like it."

For anyone building privacy tools, cross-border payment rails, or DeFi protocols with significant Iranian user base - this is your heads-up that compliance theater might not be optional much longer.

Seedance now supports 16:9 → 9:16 aspect ratio conversion via single prompt. This means you can batch-convert landscape videos into vertical format for IG Reels, YouTube Shorts, and TikTok without manual cropping or editing.

The real value: automated content repurposing pipeline. Instead of shooting twice or using third-party editors, you feed horizontal footage and get platform-optimized verticals out.

This is huge for creators with existing horizontal libraries — suddenly all that content becomes reusable across short-form platforms. The prompt-based approach suggests the model intelligently reframes (likely tracking subjects and recomposing shots) rather than just center-cropping.

If you've been sitting on a backlog of 16:9 content, this unlocks instant multi-platform distribution.

China's AI hardware scene is pivoting hard: compact models + aggressive post-training optimization + direct hardware deployment. The stack is getting vertically integrated—train small, tune aggressively, ship on custom silicon. This mirrors edge AI trends but with way more emphasis on inference efficiency and local compute. Basically ditching the cloud-first paradigm for "train once, deploy everywhere on specialized chips." Makes sense given data sovereignty concerns and latency requirements. Expect more ASICs optimized for 1-7B parameter models with baked-in quantization support.

Chinese tech giants ($BABA, Tencent, Bytedance) are flooding the market with free AI integrations everywhere. It's a land-grab play—embed models into every product surface to lock in users before monetization kicks in. Think Douyin AI assistants, WeChat AI features, Alibaba Cloud inference APIs at near-zero cost. The strategy mirrors early mobile internet: distribute first, extract value later once you own the behavioral data and usage patterns. This creates massive inference infrastructure demand and trains models on real-world interaction data at scale. Western companies charge per token while Chinese players are subsidizing adoption to build moats. Different playbook, same goal: whoever owns the AI touchpoints owns the next platform layer.

Field Application Engineers are basically the FDEs (Field Design Engineers) but for manufacturing instead of software. They're the technical bridge between product design teams and actual production lines—debugging real-world manufacturing issues, optimizing tooling setups, and translating engineering specs into actionable factory floor instructions. Think of them as the people who make sure what works in CAD actually works when you're stamping metal or assembling PCBs at scale. They deal with tolerance stack-ups, process variation, and equipment quirks that never show up in simulations. Essential role that doesn't get enough credit in hardware development.

Physical AI development diverges heavily between China and the US. US focuses on foundation models and software-first approaches—companies like OpenAI and Anthropic push reasoning capabilities before deployment. China goes hardware-heavy immediately: robotics factories, autonomous vehicles, and manufacturing automation at scale. The regulatory gap matters too. China's government accelerates physical deployment through industrial policy and less liability friction. US deals with safety reviews, liability concerns, and slower infrastructure adoption. Compute infrastructure also splits: US leads in training clusters (H100s, custom TPUs), while China optimizes inference deployment across edge devices and local manufacturing. The real kicker? China's physical AI gets real-world data faster through mass deployment in controlled environments. US builds better models in theory, but China stress-tests them in factories and streets first. Different philosophies: US wants AGI that thinks, China wants AI that moves and produces.