Meanwhile, check out this killer drone! #
Tuesday. The Q2 technology earnings cycle truly gets underway tomorrow. Alphabet will disclose its second-quarter results, alongside Tesla, IBM, AT&T, and ServiceNow. Thursday brings earnings from Intel, T-Mobile, SAP, Comcast, Nokia, Infosys, AppFolio, and RingCentral. We’ll have our eyes out for notes on compute constraints (or a lack thereof), and software health more generally.
Today, we’re looking at big drones, Ramp’s latest adventure, why we were too mean to Databricks, and notes on why you can’t stop open AI, which has major implications for closed-source darlings. To work! — Alex
📈 Trending Up #
Chinese compute capacity…AI on the loose…government AI chops…chip prices…DoD AI cost control…SpaceX earnings on August 4th…Shanghai GDP… public debt inEurope, theUnited States…collective spectrum auctions…
**Huge Autonomous VTOL Drones of Death: **A good band name, and also a new product from Anduril, the Silicon Valley defense prime. Dubbed Thunder, the twin-rotor aircraft can launch straight up and fly in a straight line thanks to its angle-changing propulsion (think V-22 Osprey). Even better, it has more “stowed kills per sortie than [an] Apache [attack helicopter],” per its creators. It looks cool to boot.
- I wonder if we’ll wind up in more wars as a species when the quantum of force becomes an autonomous piece of technology instead of human flesh and blood. Not that humans won’t get hurt; it just may be less tricky to start a conflict when your main weapon isn’t predicated on risking your own team’s lives. (Another thought: National wealth, not manpower, could decide future military power to a degree that we haven’t seen before.)
- Recent successes by the autonomous drone-boat company Saronicmatch this theme. - Recall that defense startups building autonomous weapons
are a global project, and thus one from which we should expect to see globally distributed fruits. **Doing more: **Ramp, a massive private company with a history of quick product releases, is staying true to form. Its latest product, however, is not a new corporate card or method to more quickly ingest and comb through expense reports. Instead, Ramp announced an AI model router, a tool it uses internally to better match tasks to models. As AI inference becomes a real cost center for companies, ensuring they aren’t driving a Bugatti to the supermarket (using Fable 5 to summarize an email) has become critical work.
Ramp claims that its ‘Router’ product — deft name there, Ramp — can drive a 30% reduction in AI spend at the cost of just 30 milliseconds of latency; for nearly every AI use case, that’s a fine tradeoff.
I’ve been floored this year by the pace of software development. It seems that the companies closest to the AI epicenter are building more, faster. A prime example of this is Anthropic’s aggressive rollout schedule, which in June released Claude Fable 5, Claude Mythos 5, Claude Tag, Claude Science, and Claude Sonnet 5. May brought agents for financial services, Claude for Small Business, and Opus 4.8. April saw Claude Opus 4.7 drop, alongside Claude Design, updated election protections, and Claude for Creative Work. In a single quarter’s time, the AI lab released a blizzard of new models and products that greatly expanded its work horizon. (Remember yearly release cadences? Hilarious!)
The promise of much of the reorganization of technology companies towards smaller, AI-enabled teams was speed. I think we’re starting to see the fruits of those workflow tweaks. You might file products like Oz from Warp, New Relic building an agentic platform, or Runway ML launching Dev under the *‘quickly moving tech companies building remit-extending features’ *heading, alongside Ramp Router.
- Adding to the point, Ramp CEO Eric Glyman jokingly tweeted“ramp web services 🤔” as the opening line of his post announcing Router. Go for it, Eric. - Our
thoughts on Gusto and the rise of general-purpose agents fit here, too. **Apologizing: **I owe Databricks an apology. I mocked the company’s latest fundraise as another cowardly act to avoid going public. That was incorrect, as it turns out. (Databricks should still go public, but let’s leave that particular beaten horse alone.)
Per a CNBC interview discussing his company’s latest fundraise ($3 billion at a $188 billion valuation), Databricks CEO Ali Ghodsi said several interesting things: Agents use a lot of Databricks, which charges on a consumption basis. So, as use of agents rises, so too does growth at Databricks. Ghodsi claims his company has seen “acceleration of revenue in every region, Europe, Asia, Americas, and every product line that [it has for] six or seven quarters” thanks to agentic use. The CEO also said that his company’s Unity ‘AI gateway,’ which can be used to control AI costs, has “exploded” in demand terms.
Most importantly regarding its latest fundraising, Ghodsi said that the cash has a very specific purpose (transcript tidied for readability):
[W]e are hosting these open source models like Kimi [K3]. We actually offer it to our customers and we’re running out of GPUs everywhere. We ran out of GPUs [in] Asia. Every country wants these GPUs: Japan. Korea, United States, India. So we just needed to go get a lot of these and that requires a lot of funding. That’s actually what triggered this fundraise for us, because we just were inundated with these requests. We needed to get GPUs. GPUs are costly to get. People are expecting that you’re gonna pay for [the chips]. So that’s really what triggered the fundraise in the first place. But then, of course, Kimi 3 [was] released, [which] I think is a game changer.
Fine. Raising more capital to buy the GPUs it needs to serve its customers is a fine reason to do so. Not that Databricks couldn’t have raised capital in an IPO, but I promised to hush about that.
- For more on Databricks’ revenue
acceleration,head here. Codex: After bundling ChatGPT with its developer tool Codex, OpenAI has seen quick uptake of the product. From one million active users in February, to five million in June, to now 10 million, per Bloomberg. While the world is salivating over new open-weight models, keep in mind that the closed-source labs are hardly drying on the vine.
[📉](https://finance.yahoo.com/news/servicenow-pledges-1-5bn-investment-110000403.html) Trending Down
[📉](https://finance.yahoo.com/news/servicenow-pledges-1-5bn-investment-110000403.html)
Free speech…coal…AI music…humans using the Internet…free trade…hopes for a ceasefire? …Indian GDP growth? …banking regulation…
Down the weights? The FT reports that the Chinese Ministry of Commerce is “considering tightening export controls on artificial intelligence and semiconductor technologies,” including the possible end of releasing model weights from AI models built in the country; open-weight AI models from China have found wide uptake globally as lower-cost alternatives to closed-source, and expensive American AI technology.
The FT adds that “China would still let overseas customers access the models and services, however.” This is not the first time there have been rumblings about AI restrictions from China, most notably this Reuters report from early July. In contrast, CCP premier Xi Jinping recently argued that AI development should “adhere to the principle of openness and win-win and boost innovation-driven development,” per an official transcript of his speech.
How to square the two stories? By allowing Chinese companies to build and sell AI models, albeit in a non-open format. That would allow the country to continue scaling its AI footprint without effectively subsidizing American companies building on top of open-weight Chinese IP. Let’s hope that these ideas don’t make it to market.
- How big a disappointment would it be if Moonshot’s Kimi K3 model didn’t meet its stated July 27th target for releasing its weights?
Our promised open AI future #
Today, let’s be declarative: No government can halt open-weight and open-source AI development; no government can block open-weight and open-source distribution; therefore, no government can stop use of open AI models.
That’s true simply because building software is permissionless, and even attempted government controls won’t stop open AI models from being created, shared, and used. If *North Korea *can’t keep South Korean media out, what chance would the US government have at blocking the use of Chinese AI models internally? The thought is risible.
Given that we will always have access to performant, open AI models, where does that leave us? In a price war, something venture investor Bill Gurley discussed in a recent WaPo OpEd:
A market growing fast, with rich potential margins and a long runway, is the most powerful lure in capitalism — a signal, broadcast to every capable engineer and every investor on the planet, that there is enormous money to be made by offering the same thing for less. Jeff Bezos […] built a
[trillion-dollar company]on[the principle]that “your margin is my opportunity.” By that logic, a business promising the highest returns in software summons the largest crowd of competitors in software. The open-model wave is not an attack on AI companies. It is the market responding to the fortune they say they are about to make. They called it forth themselves.
I think we’d all like to live in a time where AI is cheaper. Gurley also argues that open models enable companies to tune them with internal, proprietary data without concern (true), and that they are inherently more secure than their closed rivals (more true in an open-source context than open-weight). All told, the open AI future appears shiny indeed.
- Who doesn’t want to pay less to get more? Conversely, if you have to pay more to get less,
your entire economy could drag. Not so much for OpenAI and Anthropic investors, who want to see their enormous investments into the two labs come good, and quickly. But we are not in the business of protecting incumbents; if something better comes along, down with the king.
If the market moves away from closed-source AI models, there will be damage. But there will also be new winners. Another venture investor, Josh Wolfe, argues that China is building open AI models because it has lots of technical talent and because doing so allows it to ‘dump’ and ‘competitively undermine’ US investment in AI. But as Gurley notes in his piece — and we have written repeatedly — open AI is not merely a Chinese game. Mistral makes open models, yes, but so does Nvidia, Thinking Machines Lab, and Reflection AI. Caveats: There are three. While it might be easy to declare closed-source AI labs dinosaurs, and begin fine-tuning our own models in earnest, consider:
- We don’t know how much Chinese AI model intelligence is distilled from closed-source American frontier technology. The answer could be zero today; it has not been zero historically. Therefore, if we move inference demand from closed models to open rivals, we could starve the genesis point of AI intelligence gains.
- As we saw above, China could limit the release of model weights, turning its own AI industry into a lower-cost, but similarly closed-source setup to what we see from Anthropic and OpenAI. If that happened, ironically, American and European AI companies would find themselves at the forefront of open AI, and, in that scenario, closed-source AI to boot.
- If we shift AI revenue from today’s frontier labs to open competitors, we could lose the next Mythos-style edge over other countries. Yes, building Anthropic and OpenAI has been incredibly expensive, but they have kept American technology and industry at the very tip of the technological arrowhead in terms of performance thus far.
But as we saw with Mythos, simply having a great model today is not sufficient; you have to get the government’s permission to release it. So, the gains from a theoretically far-smarter model than other nations possess could prove more limited than we may have expected in a less regulation-heavy world.
One more point. Open AI models dissolve regulatory teeth. When open models are performant and available, what could regulation look like? The government might say ‘no using these models if you want to work with us,’ which could have some impact. But Anthropic partners are still selling its models, even as the company wrangles with the government declaring it persona non grata. And can the government really audit every partner down to the model-router level? Hell, model routers themselves disintermediate model selection, right?
Some of the folks in favor of incredibly limited AI regulation annoy me. I’ve probably been 10% too skeptical of the lightest of regulatory touches for AI because I find some of its loudest supporters loathsome. That’s on me.
**To sum: **No government, or collection of governments, can halt the creation and use of open-weight and open-source AI models. Therefore, closed-source AI labs will have to compete more vigorously on price tomorrow than they did yesterday thanks to recent gains in open AI model performance (Kimi K3, GLM-5.2). The trend of shifting inference loads towards open AI models could slow AI lab revenue growth and some compute buildout in the near term, but probably not over a long enough time horizon, as cheaper inference will drive increased demand. When AI models are open and freely available, regulation is a moot point. Thus, we’re heading for a world where much, if not most, inference will be served cheaply by model companies and cloud providers competing more on price than performance, and we all get more for our dollar over time.
You don’t have to twist my arm.