Signal #71: Albania's AI Minister Turned One and Bought Nothing

A year ago Albania gave an AI system a ministerial title and the story went around the world twice. This week BIRN went and checked what it has actually bought, and the answer is a draft document.

That distance between the announcement and the installation runs through the whole issue. OpenAI removed its risk-measurement function three weeks before its president published a warning about risk, Anthropic posted numbers nobody in this industry has posted before while saying very little about them, and the most expensive piece of AI infrastructure sold this week is one most people could not have named on Friday.

Diella was appointed Minister of State for AI in September 2025 to take corruption out of public procurement. Nearly a year later, BIRN went to check what exists and found the Public Procurement Agency has only just finished drafting the terms of reference, the document you need before you can tender the system that would do the work. Agency director Reida Kashta put it plainly: there is a project to develop an AI-powered procurement system, and it is still in progress. The legal basis has been sitting there since a 2024 amendment to the Public Procurement Law, and the national AI strategy and draft AI law are both still in consultation.

The announcement traveled and the installation never started. Digital policy expert Alba Brojka points out that the data protection Commissioner already lacks the capacity to monitor the platforms Albania runs today, never mind AI actors. The actor whose face and voice became the minister, Anila Bisha, had her injunction rejected and her full case opens on 22 September. Aranita Brahaj of the Albanian Institute of Science warns the country risks becoming a laboratory other people learn from.

If you sell AI into the public sector anywhere in the region, this is your calibration point. The ministerial press conference is not a buying signal and never was. The buying signal is a published terms of reference with a budget line attached, and in Albania that document is eleven months behind the announcement it was meant to implement.

KodeKloud gives you 1,280+ hands-on labs where you provision Kubernetes clusters, write Terraform configs, build CI/CD pipelines, configure Linux systems, containerize apps with Docker, automate with Ansible, and manage Git workflows. 

78+ playgrounds let you experiment freely in sandbox AWS environments, Kubernetes clusters, and CI/CD systems without risk. 

190+ courses across DevOps, Cloud, and AI pair theory with hands-on labs at every step. 

KodeKloud Engineer and 100 Day Challenges provide real-world job scenarios with automated grading that confirms your solutions work.

Stuck? The 55,000+ member Discord community connects you with peers and instructors ready to help. 

Every lab runs in a live environment. You deploy, you troubleshoot, you learn. No videos without context. No simulations. The kind of practice that actually builds confidence because you've done real work, not watched someone else do it.

Start Learning For Free

OpenAI dissolved its Preparedness team at the end of July, the group that assessed catastrophic model risk, redistributing biosecurity and cybersecurity oversight to staff on other teams. Ethics lead Chloé Bakalar and head of safety Johannes Heidecke both left. The reported framing is streamlining ahead of the IPO, in line with Sam Altman's instruction to cut side quests. On 17 August, the day that reporting landed, Greg Brockman published The Defender's Window, arguing that open-weight models with serious cyber capability are slated for release at the end of this month.

Read those two facts together and the Albania pattern repeats at a company with a trillion-dollar listing in front of it. The public argument stayed where it was; the function that produced the evidence behind it went away quietly, three weeks before anyone reported it. The pressure behind it is sitting in the Anthropic numbers further down this issue, where the listing race is now measured in quarters and a risk-measurement team does not appear anywhere on an income statement.

Brockman's own demonstration is the useful part. ChatGPT Work audited his personal site, found 13 issues in about fifteen minutes, DNS misconfiguration, an insecure jQuery version, unencrypted traffic between AWS and Cloudflare, then fixed them inside an hour. Point the same tool at your own perimeter this week rather than waiting for a vendor to sell you the finding.

Anthropic's second quarter came in above $11.5 billion, more than fourteen times the $787 million of a year earlier and double the $4.73 billion of the first quarter. Buried in it is the number that matters: positive adjusted operating income, the first any frontier lab has recorded. By the end of July the run rate had passed $65 billion, against $47 billion in May and $9 billion at the close of 2025. The confidential filing is in, and backers are pointing at an October debut at $2 trillion or more, against a last private mark of $965 billion in May.

For three years the standard defence of this industry was that profit arrives later, once the training runs stop getting bigger. The quiet tell here is the disclosure itself: Anthropic reported an actual quarter rather than an annualised run rate, which is what companies do when the actual quarter is good. OpenAI has doubled to $40 billion from $20 billion at the end of 2025, which is a remarkable year and the slower of the two.

The consequence lands on your contracts. A vendor half your stack depends on is about to acquire shareholders, quarterly reporting and a gross margin anyone can read, and the first S-1 in this category will be the first honest look anyone gets at what serving these models costs. Read it the week it drops, then ask whether you are paying the price a private company charges or the one a public company has to.

Bloomberg reported on 16 August that Stripe is closing on OpenRouter for more than $7 billion. OpenRouter raised a $113 million Series B in May at a $1.3 billion valuation, from Sequoia, Andreessen Horowitz, Menlo Ventures and Alphabet's CapitalG, which puts the step-up at roughly five times in three months. The company routes traffic for 8 million users across more than 400 models. Stripe says it does not comment on rumours or speculation, and neither side has confirmed.

Nobody held a launch event for the routing layer. It got installed one team at a time, by engineers adding multi-model fallback to cut costs and survive an outage, and somewhere in the last eighteen months that plumbing quietly became the place where model choice, price comparison and billing all meet. The company that owns internet payments noticed before the labs did. Buy the switchboard and switching models becomes a billing decision rather than an engineering one.

Go and check what share of your own model calls already leave through a gateway you do not own, because that number is usually higher than the person who owns the budget thinks it is. The thing to watch from here is whether the routing stays neutral once a payments company is setting the defaults, and whether the frontier labs answer by pushing customers back onto direct contracts. Neutrality is easy to promise while you are still the small party in the deal.

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

Productivity: Claude Code stops dying when you hit your limit. Version 2.1.234 shipped on 17 August with sessions that resume automatically when your usage limit resets, toggled in /config, plus GitLab merge request badges in the statusline and a Windows path fix that closes an NTLM credential leak. Free on existing plans. Turn the auto-resume on before your next long agent run and stop babysitting the terminal.

Design: Google switched off three Imagen 4 endpoints yesterday. As of 17 August, imagen-4.0-generate-001, -ultra- and -fast- are retired, with traffic pointed at Gemini 3.1 Flash Image. If a marketing or design pipeline still calls those model names, it broke while you were reading last week's issue. Grep your repo for imagen-4.0 and repoint it.

Marketing: a Belgrade founder is building the measurement layer for newsletters and creators. Astute launched publicly on 17 August with a $1.2 million oversubscribed pre-seed from Flyer One Ventures, Silicon Gardens, Marathon Fund and ERA. Two agents, one running partnership admin, one scanning more than a million creator posts a minute for fit and performance, with AI-search visibility tracked as its own metric.

Dev: Alibaba's newest open model fits on a single 24GB card. Qwen3.8-27B landed on 14 August under Apache 2.0 with 262,144 tokens of native context, a vision encoder, and an FP8 build, running four-bit on consumer hardware. The scores are self-reported, including 84.3% on OSWorld computer use, so treat them as the vendor's numbers. Worth an afternoon if data residency is what blocks your agent work.

Research: OpenAI is funding the people who will write Europe's AI rules. On 17 August it awarded $1 million in cash and $1 million in credits across 14 economic-policy projects picked from more than 400 proposals, including the Centre for European Policy Studies and ECIPE in Brussels. Six-month projects, results in 2027. Worth knowing whose research your next compliance obligation grows out of.

Next edition soon,

Çelik