July 2026
AI Today, AI Tomorrow
The industry landscape — and why local AI is the next big shift.
Robin Kim · Software Engineer @ OpenRouter
Intro
Who am I
- Software engineer at OpenRouter
The AI model routing layer — more on this soon.
- Previously: Coinbase
Built trading products and trading platforms.
- Founder: Gallery Labs → Moshi Cam
Mobile camera app — hit #1 in Taiwan, still live and popular across Asia today.
A quick teaser
I work at OpenRouter
One API for hundreds of AI models, across dozens of providers.
To explain why that needs to exist, I first need to show you how the AI industry fits together.
…so hold that thought. We'll come back to it.
The landscape
The AI stack: three layers
1 · Applications — what people actually touch
ChatGPTGemini appCursor
GitHub CopilotPerplexityNotion AI
MidjourneyCharacter.AIPalantir AIP
2 · Models — the intelligence itself
OpenAI · GPTAnthropic · ClaudeGoogle · Gemini
xAI · GrokMeta · Llama
DeepSeekAlibaba · Qwen
Moonshot · KimiMistral (FR)
3 · Chips & Compute — the physical machine underneath
Design: NVIDIA · AMD · Google TPU · Amazon Trainium · Apple Silicon
Fabrication: TSMC · Samsung Foundry · SK Hynix (memory)
Datacenters & clouds: AWS · Azure · GCP · CoreWeave · Together · Fireworks · Groq
open-weights
China-based
closed / US & EU
The landscape
Nobody stays in their lane
- Full-stack players: Google does chips (TPU) + models (Gemini) + apps. OpenAI now does models + apps + its own chip + Stargate datacenters.
The biggest players are vertically integrating in both directions.
- Partnerships everywhere: Palantir ⨯ NVIDIA (sovereign/government AI deployments) · SK Telecom ⨯ NVIDIA (gigawatt-scale AI cloud in Korea) · SK Telecom ⨯ OpenAI (ChatGPT for Korean subscribers)
Telcos, defense, and clouds are all wiring themselves into the stack.
- The chokepoint: TSMC. Over 90% of leading-edge chips — NVIDIA, Apple, AMD, custom silicon — are fabricated by one company in Taiwan. Capacity is sold out.
Memory (HBM, from SK Hynix & Samsung) is the other binding constraint.
- US vs China: chip export controls cut NVIDIA's China share from ~95% to near zero — and China answered with world-class open models, some trained entirely on domestic chips.
Quick explainer
What is a model, actually?
- A giant file of numbers. Billions of learned parameters — "weights."
- Training writes the weights. Done once. Costs up to hundreds of millions in compute.
- Inference runs the weights to answer your question. Happens every single query — this is the ongoing cost of AI.
- Why GPUs? A model is mostly enormous matrix math. GPUs do thousands of multiplications in parallel — and need very fast memory to keep the weights flowing.
Flavors of models
Text / LLMsImage generation
Speech & audioVideo
Multimodal — all of the above in one
Open vs closed weights
Open: the weights file is published — anyone can download and host it (Llama, DeepSeek, Qwen, Mistral).
Closed: only the maker serves it, via API (GPT, Claude, Gemini).
Metaphor: apps are storefronts, models are engines, chips & datacenters are the power plants and the grid.
The key insight
One model, many sellers
Because open models are just downloadable files, anyone with GPUs can host them — and many companies do, competing on price, speed, and reliability.
8
providers hosting the same DeepSeek model on OpenRouter
13
providers hosting the same Qwen3 235B model
~⅓
of all OpenRouter token volume runs on open-weight models
- Same model, different hosts: different hardware, different prices, different latency, different privacy policies.
- So the question stops being "which model?" and becomes "which provider of that model — right now?"
Callback
That's where OpenRouter sits
→
OpenRouter
routes each request to the best endpoint
→
400+ models · 70+ providers
OpenAI, Anthropic, Google, DeepSeek hosts, …
- Tracks uptime, latency, and price across every provider — and fails over automatically when one goes down.
- Honors your preferences: cheapest, fastest, EU-only processing, zero-data-retention-only, specific providers, price ceilings.
- One account instead of ten. Try any model without signing up for every lab and cloud separately.
Scale today: ~100 trillion tokens/month across millions of users. Everything here is from openrouter.ai's public docs.
Where it's going
The rise of local AI
Running models on hardware you control — your phone, your laptop, your company's own servers.
~6,000×
cheaper per token: electricity on your own GPU vs a frontier API
52M
monthly downloads of Ollama, the most popular local-AI runtime
93%
of surveyed enterprises repatriating AI workloads from public cloud — or evaluating it
- Cost: once you own the hardware, you basically pay for electricity. Use it as much as you want.
- Privacy & sovereignty: your data never leaves your device. No lab can train on your trade secrets.
- It works offline — and small open models today match what needed GPT-4-class cloud APIs just two years ago.
Where it's going
The future is local + cloud, not local vs cloud
- Local first for the everyday: private, instant, free at the margin.
- Cloud fallback when you need to scale, or need frontier intelligence that doesn't fit on your device yet.
- The big platforms already architected this: Apple runs a model on your iPhone and falls back to Private Cloud Compute; Windows routes between your laptop's chip and Azure.
- And the routing layer matters either way — when you do go to the cloud, something has to pick the best model and provider for the job. That's OpenRouter's role in this future.
Closing thought
Competitive markets make AI better
- Few models → little choice → pricing power concentrates in a monopoly or duopoly.
- More models — especially open ones — means better prices, more choice, and real self-sovereignty.
- Open models are already ~⅓ of the market and near frontier quality, at a fraction of the price.
- Local AI + open models + neutral routing keep that market honest.
Thank you. Questions?
Meta
How this deck was actually built
This whole presentation — research, writing, design, code — took one conversation with Claude Code.
This is the actual session that produced every slide before this one.
1 / 12
← → or space to navigate