This website uses cookies

Read our Privacy policy and Terms of use for more information.

In partnership with

Hey there dear reader

Thank you as always for sticking around and reading through the selected stories. What is interesting to point out this Friday is that something shifted in the last three days and nobody put it in a press release. The two labs at the front of this industry each published work showing their newest models doing things the previous generation could not, and each attached the same footnote: you cannot have it.

Read that next to what Jane Street paid to sit closer to the silicon, and what Washington published on Wednesday about who is allowed to sell a drone in America, and the same shape shows up three times in one week.

Money is no longer the thing that gets you through the door. Sofia, in the middle of this issue, is the one place treating that as a product decision rather than a complaint.

Claude is the best protein designer ever!

On 18 August Anthropic published results from an autonomous protein design campaign. Claude ran from a single 30,000-token protocol with no further human guidance, generated 1,320 designs against 15 targets, and got 354 confirmed binders back from wet labs at Adaptyv Bio and Twist Bioscience. The single-target hit rate was 35.1% against an industry norm of 10 to 15%. On one target, RBX1, it hit 40% where 245 human competitors in an Adaptyv Bio contest averaged 3.7%, and its top-ranked design bound more tightly than the human winner.

The model that did this is called Mythos Preview and you cannot buy it. Anthropic says protein design and dual-use biological capability "remain unavailable" for general access in its most capable model, with a verified-researcher programme planned. The same day, OpenAI disclosed that its unreleased Astra model reached the Critical cybersecurity threshold under its Preparedness Framework, paused reinforcement learning on deployment-bound models for two weeks, and left its largest planned frontier run on hold. It now runs activation classifiers at every sampled token, at a cost of roughly 20% of monitored inference compute.

The tier you can buy and the tier that exists have separated, and the gap is set by policy rather than price. Plan for capability that arrives late, and for a compute surcharge nobody wrote into your contract. Then ask your best technical hire whether your company could pass a verified-researcher review, because that review, not your credit card, is turning into the gate.

Hire Ava, the AI BDR built for enterprise

Ava is the first AI BDR to run outbound end to end, and you decide whether she runs autonomously or on copilot.

She finds leads or ingests accounts from your CRM, enriches them, sends personalized emails on behalf of your reps, follows up, handles replies and books meetings. Website visitor de-anonymization, intent signals, and a parallel dialer come built in.

A small team can manage her centrally for thousands of reps who never log in. Everything syncs two-way with Salesforce and HubSpot.

Ava runs outbound for companies like DoorDash and Grammarly, and one customer deploys her across 1,000+ reps. Ava is SOC 2 Type II audited, SSO and GDPR ready. Ava is how revenue teams grow pipeline without growing headcount.

Sofia sold the layer underneath instead

The day after Anthropic published, Oakley Capital took a majority stake in Graphwise in what Tech.eu called one of Bulgaria's largest software exits. Graphwise is the 2024 merger of Ontotext, founded in Sofia in 2000, with Vienna's Semantic Web Company: RDF knowledge graphs and semantic layers sold to over 200 blue-chip customers in finance, pharma and the public sector, growing organic ARR above 30% a year. The seller is Portfolion Capital Partners, the Hungarian fund that bought in during 2022. Terms are undisclosed.

A 26-year-old Sofia company just became the platform asset of a London buyout fund that says it will consolidate a fragmented market by acquisition. Set that beside the gating above and the logic is clean: if the best model is rented, verified and withheld, the durable asset is the governed data underneath it and the hardware it runs on. Two days earlier, eight-month-old DiscreteStack raised €800k from CleverPine Ventures for the other end of the same street: patent-pending optimisation that runs large open-weight models on a standalone server, air-gapped if the customer wants.

Founder Hristo Todorov phrased the pitch better than any policy paper this year: "European AI does not mean an American product hosted on a server in Frankfurt. It means organizations controlling the infrastructure on which their artificial intelligence runs, owning their own data, and setting their own rules." Watch whether Graphwise starts buying regional data and semantic shops over the next two quarters, because Oakley has said plainly that it plans to. Founders of those shops should expect the call.

Jane Street bought a rack, then set the price of AI silicon

Also on 18 August, Etched closed $700m at a $21bn valuation, double its July mark of $10.3bn and roughly four times the $5bn it carried in December. Jane Street led the round. Jane Street also tested the hardware and now has, in its own words, "our own rack running in our datacenter." The lead investor and the first production customer are the same firm. Etched builds one thing, chips for transformer inference: a low-voltage prefill part, plus cluster-scale shared memory for the decode phase.

A proprietary trading desk, not a hyperscaler, is now the reference customer that marks the price of AI silicon. A day later the pattern repeated in London, where Bloomberg reported that Fractile is in talks for around $600m at $6.5bn, up from roughly $1bn in May, on the back of a reported $250m chip order from Anthropic for hardware that does not ship until 2027. Neither side commented and the round has not closed. The sequence is the point: the purchase order came first and the valuation followed it.

What both deals price is access, not revenue. A signed relationship with a frontier lab or a latency-obsessed buyer is worth billions of dollars of markup before a single unit ships. For a European hardware founder that is the week's lesson: one named anchor customer moves your next round further than a year of benchmark charts. Find out who signs procurement inside the lab you are targeting, because that person is already on your cap table.

Washington made your bill of materials the border

On 19 August the Federal Register published Proclamation 11055 at 91 FR 53699. From 12:01am on 3 September, drones above 25kg, thermal-imaging drones, docking stations and a long list of critical components carry a 100% tariff entering the United States. Drones at 25kg and under carry 25% from the same date, and a further 25% hits an additional component list on 9 February 2027.

European products get a cap of 15% and British ones 10%, and this is where the sentence turns. That cap applies only where "substantially all the critical components and technology are certified by importers" to originate in the US, the EU, the UK, Japan, Korea, Taiwan, Switzerland or Liechtenstein. Being a Polish or Czech manufacturer does not earn you 15%; certifying your bill of materials to US customs does. If the flight controller, the speed controller or the camera module comes from China, the number on 3 September is 100%. A separate customs notice published the same day begins voiding inaccurate importer-of-record numbers on 18 September, and a freight forwarder's address stops counting as one.

Two weeks. That is the distance between now and the moment component provenance decides whether a European drone company has a US market. The work is unglamorous and it starts today: pull the bill of materials, mark every part by country of origin, and find the share you can certify without argument. Companies already on the Pentagon's Blue UAS list or the FCC conditional approval list get pushed to roughly 9 February, the strongest argument anyone has made this year for paying what that qualification costs.

Short Signals

Five things to install or test this week, tagged by the seat they help.

Dev: a coding agent that fits in six megabytes. Vercel Labs open-sourced fx, a CLI agent harness written in Zig under Apache 2.0, with v0.0.4 landing on 19 August at 6.39MB. Model-agnostic across local models, gateways and subscriptions, with sessions, sandboxed commands, MCP servers, skills, subagents and a headless mode. It is marked experimental. Install with curl -fsSL https://fx.sh/setup.sh | bash and point it at a repo you do not mind breaking.

Research: Grok 4.6 arrived on Bedrock with a residency switch. Generally available since 19 August at $2 per million input tokens and $6 per million output, with a 500K context window and four reasoning-effort levels. The part that matters for regulated work is the pair of inference profiles: us.xai.grok-4.6 pins routing to US regions, global.xai.grok-4.6 does not. If a compliance review has been blocking a model choice, this is the cheapest way to unblock it.

Design: QA agents that crawl the site before your client does. Superflow AI launched on 18 August with agents that review a site and pin dead links, typos, broken forms, accessibility failures and brand-guideline violations as comments a human approves or rejects. Four agents ship free (accessibility, links, spelling, OG images), it installs on Webflow, Framer, HubSpot, Shopify and seven more, and the first 50 reviews cost nothing. Upload your launch checklist and point it at one staging page.

Marketing: ChatGPT ads opened across 31 European markets. OpenAI announced the expansion on 18 August, starting with its Ads Solutions team and agency partners, with self-service following later in the summer. Conversion optimisation, geo-targeting and custom audiences are included. Inventory sits on Free and Go plans only, so Plus, Pro and Enterprise stay clean. New-market inventory is the cheapest it will ever be right now, so run a small geo test before the agencies arrive.

Productivity: a free year of Google's paid tier for every student on your team. From 19 August, students in more than 140 markets get Google AI Plus free for twelve months, normally $4.99 a month, with Gemini Omni, doubled usage limits and 400GB of storage. US students get AI Pro instead, normally $19.99. Albania sits on the excluded list alongside the US, Canada, Bolivia, Hong Kong, Macau and Tunisia. The deadline to claim is 31 December.

Have a great weekend and hope you’re enjoying digjitale newsletter. If you do, share it with your friends and help me grow it.

Çelik