🕶️ Old GPUs just embarrassed Cerebras

Good morning, future legends.

The most expensive question in tech right now isn't what to build. It's what to squeeze harder. This week, the smartest bets are being placed on optimization over acquisition, wringing speed from old hardware, wringing hours back from agents, wringing focus from saying no.

Meanwhile, billion-dollar infrastructure plays are getting repackaged as tradable securities, and platforms are learning the hard way that friction has a legal price. What gets leveraged next matters more than what gets funded.

The established tenet that AI inference bottlenecks demand custom silicon just got mugged in a French alley. Turns out the biggest gains in AI acceleration aren't hiding inside next-gen chips. They're already inside the GPUs you own, waiting for someone insane enough to write code at the binary level.

Meet Kog, an 11-person French startup that treats GPU optimization like a hacking challenge. Fitting, since CEO Gaël Delalleau is a four-time DEFCON CTF finalist with a solid-state physics background. He isn't tuning knobs. He's rewriting inference at assembly and binary levels, the digital equivalent of picking a lock with a scalpel.

Here's the flex. Last May, Kog ran a 2-billion-parameter model on standard datacenter GPUs (AMD MI300X and Nvidia H200) and hit 3,000 tokens per second, claiming 30x faster inference. That demo generated 200 business leads and a seed round co-led by Varsity VC, plus Bpifrance and Scaleway.

Meanwhile, on the silicon layer… Cerebras IPO'd this May betting billions that only bespoke chips win. Kog says the cheap hardware already works. You just have to torture it correctly.

Scaling from 2B toys to real LLMs takes weeks per GPU type. The whole thesis rides on hitting 10x speed on a first major customer model by September 2026. Nail that, and the Series A writes itself.

🥨 Gheelish secures shelf space in 2,000+ national retailers after Darin Alpert's cold-email blitz wins buyers.

Instead of pitching category buyers with a generic deck, he sent a tightly personalized three-touch sequence that opened with their exact planogram gaps, followed with velocity data from early Sprouts turns, and closed by proposing pre-funded promo calendars and in-store sampling support. That operational readiness made it easier for chains to greenlight multi-region tests that quickly converted into nationwide authorizations.

🤖 Google replaces classic search links with AI answers, triggers 60% traffic drop for small publishers, and debuts first official AI optimization guide.

Behind the scenes, budgets are already moving: PMG is steering brands to pour 1.5–2x their current search spend into generative engine optimization pilots, and a June 2026 Semrush survey shows 38% of marketers now prioritize AI search optimization versus 36% for classic SEO, with 85% reporting AI has changed their search playbook.

🤝 72% of co-fouwnder breakups at U.S. startups stem from outdated, unmaintained agreements, not headline disputes.

The tension usually surfaces when a departing founder’s oral promises collide with a cap table and vesting stack that no longer match reality, forcing emergency “cleanup” rounds of retroactive stock repurchases, IP assignments, and revised board control. That scramble often delays fundraising by quarters, drives up legal spend, and permanently locks in awkward ownership for whoever stays.

🔍 Rich Barton transforms Expedia, Zillow, and Glassdoor by unlocking insider data for 500M+ users worldwide.

His playbook starts with a deceptively simple prompt: “what do insiders know that outsiders would pay to see?”, then scales it by productizing that hidden data into consumer-facing price comparisons, home valuations, and salary benchmarks. Once those data flows reach critical mass, the platforms flip into two-sided marketplaces where advertisers, employers, and agents fund the free tools that keep users locked in.