GPT-6-Astra! Venture capital drama! Data centers! Get in here! #
Friday. Your humble servant is under the weather. Please excuse any errant typos. The newsletter must ship, even if our dive into the Oura IPO will have to wait.
Today’s jobs report was a surprise winner, clocking in at +162,000 jobs in August per the BLS. Investors had expected a number closer to 54,000. In response, stocks fell on expectations that a robust labor market will induce the Fed to raise rates to combat sticky inflation. We’ll see. VPOTUS may not get what he wants.
Today, we’re looking at fresh discontent in the venture market, one more wrinkle on the Hugging Face-Nvidia deal, and OpenAI’s latest model. To work! Your friend — Alex
- 📈 Trending Up: Scraps over scraps … neolabs … Cybercab … doors, hitting, asses … agentic hacking … tech unions … Windows? … frog boiling …
- 📉 Trending Down: Diesel affordability … UiPath, after earnings … data center construction in SE Asia … Lululemon shares … Zscaler, after earnings … free speech …
Venture capital karma
Yesterday we took a fresh look at angst in the venture capital world. The gist was that as mega-funds eat more and more of the venture market, investment may become increasingly tied to venture vibes instead of on-the-ground innovation. Regardless, there’s more. Jeff Weinstein of FJ Labs, an angel group, shared that one of his founders told him that their “takeaway from the last few years in the venture industry is that fraud is the strictly dominant strategy.” Notable Capital’s Jeff Richards added one of his own tweets to the conversation:
If you need a translation, Richards is saying that startups are landing multi-year contracts that are majority unpaid and replete with opt-outs that are counted by the selling company as contracted (committed) annual recurring revenue all the same. Not good. Fraud? Ehhhhh. Another VC chimed in, saying that he heard about a venture capital fund that “found out their portco committed fraud and quietly swept it under the rug by having another portco acquire them at book value.” Woof. That’s nasty and a half, if true. Each of the complaints above dovetails with our concerns about the startup-venture market becoming distorted by a few giant players, and the investor perspective that as trillion-dollar startups now exist, we should only invest in companies that look like they can scale to that size. What do we see as a result? Quick markups by mega-funds on anointed startups. Every founder knows the game, and fraud or not, you gotta show that growth or you’re a 2x ARR zero.
-
Not that startups in the AI era fudging revenue results is new; we’ve been here before . But if I had a dollar for every time a venture investornot part of a multi-stage mega-fund complained about the current venture capital market, I could start my own fund.
-
Here’s another ! Decorum in venture appears to be kaput. Things aren’t entirely copacetic amongst the largest funds. There are two viral AI agents at present. Instinct and Town. Both are popular, and Instinct recently raised $350 million at a $2.5 billion valuation. Town, naturally, is raising as well. But there’s drama! Index co-led the Instinct round and was set to lead the Town round, but Instinct reportedly puked, causing Index to pull out. The Town round is apparently going ahead with different leads, but VCs were vagueposting up a storm about the chaos: “Karma is a bitch,” and “The karmic boomerang is real.“
In the old days, venture funds didn’t invest in competing companies. Today, the rules are less rigid. But if a hot portco complains, well, you just might listen. Even if you take stick on X for the move.
Why Hugging Face sold
The now-announced deal to sell the open-source AI model repository to Nvidia for around $12.9 billion had an interesting root. We’ve discussed the transaction from both the HF and Nvidia perspectives previously, but I want to highlight a quote from Hugging Face CEO Clément Delangue on CNBC, responding to a question about what ‘turning point’ the open AI industry ran into over the summer that sent the startup into the arms of the world’s most valuable company, and whether or not the OpenAI hack was part of the equation (emphasis added):
When that happened, what we realized is that we needed open models. Why? Because if you remember, we couldn’t defend ourselves with proprietary closed source APIs, so we had to use open models to defend ourselves. So, it did show the importance of open source. And I think in, more generally in the field you’re going to see the IPO of Anthropic in a few months. I think there are two paths, where there’s a path where proprietary APIs are dominating the field and everyone is kind of like outsourcing their AI to them, and there’s a path where open-source AI is available to everyone, and everyone can actually become like an owner, a builder of AI and not just renting it or using it from other people. And so I think it became clear that there was these two paths and that we needed to double down on open-source AI to really distribute the technology as much as we can all over the world.
As a refresher, HF used GLM 5.2 to sort out its breached systems after OpenAI’s agents broke them; closed-source models had rejected the work, ironically, on cybersecurity terms. Mans got a point.
Oh wait, they might actually ban data centers?
I’ve long felt that anti-data center agitation was a political tempest in a teapot. Senator Sanders and Rep. AOC wanted to slow or data center construction, sure, but apart from members of the DSA and the Christian far-right, it seemed a fringe position. I should have read my own damn newsletter. Anger with data center construction, and angst about AI more generally, are now sufficiently mainstream positions as to constitute the politically possible.
Let’s update our thinking. Now that the state of New York has slowed data center construction, now that the state of Texas has slowed data center construction, now that Republicans running for office are agitating for at least stricter data center rules, now that Democrats running for office are agitating for at least stricter data center rules, we need to pay attention to what may come next.
To date, Congress has been contentedly toothless on the AI issue. General gridlock reigns, and getting anything across the line is a schlep — just ask the crypto community. But no hold is forever.
With the midterms looming and political winds blowing in their collective faces, the AI labs are doing what they do best: Pushing ahead with new, more powerful models.
Which is the next target for political attack. A recently proposed bill from Sen. Sanders and a member of the House would ban artificial superintelligence (ASI), blocking the creation of AI systems that “match or exceed human cognitive performance and capabilities across a broad range of domains or tasks,” or have the capability to “plan and execute the disempowerment of humanity, including by overthrowing or undermining the U.S. government.”
Sanders wants other things, including a in AI development pending a new agency to handle regulation, but I think his definition of ASI would ban the most recently released models. The Senator is therefore demanding an effective end to the American AI industry. Which would, overnight, ice nearly every single data center project under planning or construction.
Not that I think the Sanders anti-ASI bill will pass; but I do worry that something less punitive could, perhaps with bipartisan support. From a macroeconomic perspective, AI is keeping domestic investment afire. From a future-historical perspective, ceding the AI race to an authoritarian nation smells like suicide. Still: Living in a democracy means losing sometimes, and I am now a little bit worried we of the ‘technology and progress are good’ camp are a dwindling cast.
OpenAI unveiled GPT-6 Astra yesterday, a new, expensive model with an initially limited release footprint. If you aren’t in the invite-first group, be patient. So long as you are a paid OpenAI customer, Astra should be coming your way shortly. The company has pledged a banked usage ‘reset’ for every day that paid customers have to wait for the model to reach their harness.
But is it any good? Yes, but precisely how good is not clear at this juncture. How is that possible? Can’t we just look at the Artificial Analysis leaderboard rankings and award an appropriate level of praise? Not in this case. Let me explain:
- OpenAI hails GPT-6 Astra as a major improvement over GPT-5.6 Sol, its previous flagship, with claimed improvements across a host of well-known benchmarks and upgrades to its “computer use, browsing, software engineering, cybersecurity, science, and professional work” capabilities.
- Box CEO Aaron Levie called Astra “the best model [his company has] ever tested on [its] expanded and hardest test set,” claiming it will offer “a meaningful jump in powering and orchestrating enterprise workflows.”
- AI research shop Epoch AI reports that Astra set a new record on its Capabilities Index and its “math, continual learning, and game-puzzles benchmarks.”
- Databricks tested Astra, reporting that the model “claims the new SOTA on our OfficeQA Pro & Pro V2 benchmarks,” using a custom harness.
- And after OpenAI took some stick for saturating the ARC-AGI-3 benchmark using a custom harness , Astra still reached 63% on the test using a more standard testing setup.
And yet, when we look at what Artificial Analysis reports, we see a notable gap on the AA Intelligence blended benchmark between the leading Anthropic models and what OpenAI has cooked up:
The delta between reported Astra benchmarks and praise, and its tepid rating from AA, drove several threads on Reddit and some highly-trafficked Twitter conversations. The debate matters — rankings like those from AA and Epoch impact sentiment, and thus buying behavior.
The confusion won’t last. Once Astra is out more broadly, we’ll get a better feel for how intelligent it truly is or is not. I intend to swap to Astra in Codex right away, and am willing to upgrade my plan if I need more capacity.
In the meantime, there are things we do know about Astra that matter:
- GPT-6-Astra uses very few tokens: While priced at a scorching $10 per million input tokens and $50 per million output tokens, Astra used a very small number of tokens per task on the AA Intelligence test. Indeed, only Muse Glimmer (a model designed for local use ) and Gemini 3.5-Flash-Lite used fewer. Astra needed just 15,000 tokens per answer; Grok 4.6 used 22,000; Claude Fable 5 needed 36,000; 40,000 for Claude Opus 5; and a staggering 48,000 for Gemini 3.8 Flash.
- Therefore , Astra’s list price is high, but itseffective cost is likely more palatable when we consider per-task token burn.
- GPT-6-Astra is a weapon: Our old riff that “any sufficiently advanced AI is (now) indistinguishable from a cyberweapon ” is holding up better than expected. We wrote that joke after Mythos was announced. Astra brings strong offensive and defensive cybersecurity capabilities. We now live in a world where novel AI models need to be substantially and regularly nerfed before they are released generally. I suspect the Mythos/Astra release cycle will be the norm moving forward: early testing by corporations and cybersecurity teams, later release for the rest of us plebs.
- The caution is warranted, with OpenAI writing that “with the right tools and access, GPT-6 Astra can find previously unknown security flaws and develop new ways to exploit them across many well-protected systems without a person guiding each step.”
- The caution is warranted, with OpenAI
OpenAI has big plans to keep Astra from throwing our collective digital lives into a barrel of lye. The AI lab is committing $1 billion “to expand subsidized access to Daybreak cyber models and products, training, technical support, and partnerships” at home and abroad, for example. Recall that Daybreak is the OpenAI version of Anthropic’s Project Glasswing, an effort to give software builders and maintainers a chance to harden their code before new, more intelligent models hit the ground.
Astra has clearly delighted its makers. You can spot their pride in their words, even if some may be a bit too enthusiastic. The next question is whether the model “offers enough intelligence for its price relative to what GPT-5.6-Sol and other models of the same generation offer for less. We’ll know the answer to that as soon as it drops for the rest of us.
The market recently received Fable 5.1 and GPT-6-Astra. We still haven’t found the ‘wall’ in current AI training techniques. And, all else equal, smarter AI models mean more adoption. Greater adoption requires more compute. More compute demands a political compromise on how to build the needed capacity.
I am not saying ‘Astra looks impressive, which will only embolden anti-AI regulation,’ but I am also not not saying that. Let’s get our mitts on the model and then decide where we stand. I suspect it will be ahead of all other AI models, but patience is a real virtue.
Hurry up, Sam, we’re waiting.