TechLook
Today's signal is pretty simple: the AI race is shifting from model bragging rights to control over the whole stack. OpenAI wants its own silicon, Google is pushing models into browsers and desktops, and the competitive moat is starting to look more like infrastructure than chat quality.
OpenAI tries to own inference economics
OpenAI's custom chip push is the biggest structural story in today's file. The company is no longer just optimizing models; it's trying to own the hardware economics underneath them.
Here's everything you need to know:
- OpenAI says its first custom ASIC, Jalapeño, was co-built with Broadcom in nine months.
- The company says that timeline is about half the usual 18-24 month ASIC cycle.
- Jalapeño is already running GPT-5.3-Codex-Spark at target power.
- OpenAI says the chip cuts inference cost by 50% per token versus current Nvidia GPUs.
- The chip is for inference, not training.
- OpenAI is targeting 10 GW of compute powered by custom chips by 2029.
- Nvidia still anchors OpenAI's model training.
OpenAI is telling the market it doesn't want to rent every layer of intelligence forever. If the cost savings hold, this is the kind of move that changes pricing, margins, and negotiating power all at once. It also sends a message to every serious AI lab: the real competition is now compute control, not just benchmark scores. The open question is whether this stays a one-off success or becomes the template for the next wave of model companies.
Google pushes AI agents into the desktop
Google DeepMind is baking computer-use directly into Gemini 3.5 Flash, which means agents can start operating across browsers, apps, and desktops without a separate orchestration layer.
Here's everything you need to know:
- Gemini 3.5 Flash now includes native computer-use functionality.
- The model can navigate browsers, apps, and desktops.
- This removes the need for separate computer-use tooling in some workflows.
- Google is positioning the model for agentic automation, not just chat.
- The change makes Gemini more useful inside real work environments.
This is the kind of feature that quietly becomes a distribution advantage. If Google can make agents feel native inside the desktop, it lowers the barrier to real workflow automation. Founders should care because this is where model capability turns into product stickiness. The battle is moving from model APIs to who owns the interaction surface.
Anthropic says the model theft fight is real
Anthropic's claims about Alibaba are another reminder that the AI race is not just about building better models. It's also about protecting the work that went into them.
Here's everything you need to know:
- Anthropic told the US Senate Banking Committee that Alibaba ran about 25,000 fake accounts.
- Those accounts sent 28.8 million queries between April 22 and June 5.
- Anthropic says the goal was to copy Claude outputs.
- Alibaba stock reportedly fell to a 16-month low near $99.
- Anthropic previously flagged DeepSeek for similar behavior.
If this is accurate, then model distillation has become a geopolitical and commercial weapon, not just a technical trick. That matters because frontier labs now have to defend the answers, not just the weights. For founders, the takeaway is blunt: the value in AI is increasingly tied to what can be protected, audited, and monetized, not just what can be demonstrated.
ByteDance shows production video is getting closer
ByteDance's Seedance 2.5 matters because it is pushing text-to-video closer to something teams could actually build on, not just demo.
Here's everything you need to know:
- Seedance 2.5 can generate a native 30-second 4K clip from one prompt.
- It supports 10-bit color.
- Audio is generated in the same latent space as visuals.
- The model accepts up to 50 multimodal reference inputs.
- It includes a 3D white-box preview for rough animation.
- ByteDance claims 20% better prompt adherence than the prior version.
- Enterprise beta is live, with public launch targeted for early July 2026.
The important part is not the wow factor; it's the production path. If the output is stable enough, this stops being a toy and starts becoming a workflow tool for marketers, creators, and ad teams. The reference-input limit alone is a signal that video models are becoming more controllable, which is the thing builders actually need. The gap between demo and usable asset is shrinking faster than most people expect.
⚡ Quick Hits
- Qualcomm: Qualcomm is moving into data-center CPUs with Meta as its first customer, which is another sign that AI infrastructure is attracting every major chip player.
- Micron: Micron is benefiting from the AI memory crunch as DRAM and HBM supply stays tight, so the bottleneck story is spreading beyond GPUs.
- Krea: Krea 2 is open-sourced with claims of native 2K image generation in about two seconds, which raises the floor for fast creative tooling.
- Fable: Signs of a possible comeback are emerging after the compliance freeze, but it is still tied to legal and political uncertainty.
- Mercury: Mercury Command lets users say things like 'pay this invoice' and have the AI execute inside the account with approval steps.
Techlook — AI & tech signal for founders and builders.