Grand strategic reflections and ramblings
Dustin’s raw biz-dev-strategy brainstorm. Riffing, not conclusions — the open questions below are the point.
Questions that warrant further exploration / discussion
- Do we build the agents, or do others? Do we need to build all the agents in our network, or will other people build agents that eventually tap into our network? (flagged to revisit)
- Will the labs encroach? Will the labs move into the secure/reliable/trustworthy inter-agent information-exchange territory that we see as our core? (flagged to revisit)
- Would Palantir and/or the labs facilitate cross-entity comms? Validate the thesis that Palantir won’t (a leak would undermine their core business). Want Bliss’s thoughts on Palantir and Max’s thoughts on OpenAI.
- Why hasn’t Palantir opened its infrastructure? Why hasn’t Palantir done a labs-style move — licensing Gotham as infra for others to build on — to entrench itself in the data-infra stack?
- What about the labs specifically? Would they move into cross-entity agent comms, and how reluctant would they be initially given the trust/security risk to their core application-layer opportunity? (discuss w/ Bliss)
- Multi-model / open-weight world: How would a Codex agent communicate with a Claude agent? If the world doesn’t trend toward an LLM monopoly — and open-weight models (Chinese ecosystem, NVIDIA’s new initiative) proliferate — who/what facilitates trusted comms between these agents?
- Direct comms vs. validated comms: Are we sure we don’t want to review/validate the comms, and instead let agents communicate directly? Is the ideal product a plain protocol or a trusted protocol with leak-prevention/validation built in? Is trust the core value prop — is this almost a cybersecurity product?
- What is our actual vision? What are we really trying to build, what will the world look like, and in that world what will really be valued? Specifically: is the vision to become the foundation of the AI-hardware-services ecosystem, or the medium of trusted exchange between agents? (Different visions, same path, different hires.) Want Bliss’s long-form response.
- Are existing inter-agent protocols good/sufficient? How do you stop fundamentally non-deterministic agents from blabbing something confidential across entity boundaries? Is the answer to hire a cracked researcher from Anthropic?
- NVIDIA pipeline pacing: How much do we plow ahead on the NVIDIA pipeline (targeting the companies we want them to introduce us to) vs. wait until NVIDIA is ready and then go?
- Big dogs head-on vs. wait for intros: Do we attack AWS, Foxconn, and the rest of the chain on our own, or wait for NVIDIA to make the intros?
- Start small — which small players? Who are the small players in the DC hardware space, and are they positioned for growth or to die and get eaten? Which small DC players are unlikely to get immediately squashed?
- Data centers or horizontal? Do we start in data centers at all, or look more horizontally? If horizontal, which verticals (including medical)?
Full brainstorm
So we need a biz dev strategy. The same way I pushed Bliss to develop a product vision and product strategy, we need something on the business side.
What does that look like. Well, the way I view it there are two paths. There’s the linear/heavy institutional path And there’s the scrappy path
Let’s start w what is the vision: build this network such we become the irreplaceable communication layer allowing all these different agents to communicate
So it feels like we need to start by building all these different agents so they ultimately can/do communicate Herein lies a question to revisit: do we need to build all the agents in our network, or will other people build agents that eventually tap into our network?
The core technical thing that it sounds like we can/should address is the secure/reliable/trustworthy exchange of information between agents (of different entities). Open question (another one to revisit): will the labs encroach upon this territory?
It seems like Palantir likely will not because they are focused on using their underlying software to service given organizations. The comp for us would be like Palantir opening Gotham to others to use and charging a license to others to build their own [something] on top of Gotham. Then Palantir would own the Gotham infrastructure but not the application built on top of it. Right now Palantir owns both the infrastructure and the application. Interesting. The application is a services business à la consulting. The platform/infrastructure is a product business. That’s the software. That’s the valuation.
So now let’s think about two parties as they relate to Palantir: 1) the labs and 2) project TBD.
For the labs, they very clearly own the foundational AI infrastructure. They’re trying to build/own the application layer on top of their infrastructure, but they def don’t plan to or expect to own all of it. A huge part of their business and the future is the Claude code/codex agents that other people build that sit atop their infrastructure, allowing them to monetize thru the token spend. This is a different model than Palantir chose. I wonder why Palantir hasn’t done something similar to entrench themselves in the data infra stack for these companies.
Ok so the question is whether Palantir and/or the labs would facilitate the cross entity communication that we envision. I’d love to hear Bliss’ thoughts on Palantir and Max’ thoughts on OpenAI. I don’t think Palantir would do so because their entire business is predicated on using their infrastructure to create digital twins of a specific company using their proprietary data without owning/controlling it. Any leak would make the risk of engaging with Palantir immense and undermine their core business. Therefore Palantir is unlikely to move into this cross entity communication flow. Let’s discuss w Bliss and validate this.
What about the labs? My guess is they could move into it eventually, but initially they would be very wary as their bigger opportunity is building the application layer for their products, and the biggest risk to their ability to build that application layer is people losing trust in the security of their data when engaging with Claude/ChatGPT. This is something for us to discuss. Bliss let’s chat. My sense is they could in a future state have this, but they’ll be reluctant initially. And even in the future, how would a Codex agent communicate w a Claude agent? What if the world doesn’t trend toward LLM monopoly — who/what facilitates trusted comms between these agents? And what about if/when the world moves towards open weight models? The Chinese open weight models are already kicking ass. NVIDIA just announced an initiative to build its own open weight models to counter the Chinese ecosystem. How will these agents communicate?
This, I think, we need to vet, but I think, is a big opportunity: how to become the trusted communication party/facilitator in this ecosystem. This is where I might challenge the product vision that Bliss articulated: are we sure we don’t want to review/validate the comms and instead just allow the agents to communicate directly? Isn’t, in the world outlined above, the biggest need trust? Isn’t that the risk — that one agent blabs something it shouldn’t to another? And so, is the ideal product a protocol that allows them to communicate directly or a trusted protocol with validation built in to ensure no leaks? It’s almost a cyber security product when you think about it that way. What is our core value prop? It seems like trust might be a really valuable value prop in this future world.
This is something for us to unpack — what really is our vision. What are we really trying to build. What do we really think the world will look like and, in that future world, what will really be the thing that is valued? Bliss, I’d love your long form response to this — I’m just riffing, could totally be wrong, but these are my thoughts.
And then, depending on what this looks like, we might go out and actually develop technology that allows us to get there. Like I know there are inter agent communication protocols — are they good? Are they sufficient? How do you make sure these agents don’t fuck up and blab something confidential to one another? Hard to do when you’re dealing with fundamentally non deterministic models. So how do you solve that? Is the answer that, in order to get there, we hire a cracked researcher from Anthropic? That’s exactly the kind of thing we should think thru and build into our business plan. It all stems from what do we think the world will look like, what problems will exist, and which of those problems are valuable to solve/unlikely to be solved by the labs/palantir.
And so where do we stand now. We have this deep hook inside of NVIDIA. That’s valuable. We have the ability to pursue a pure consulting playbook — build them a solution to their workflow problem, get paid, and find the next one. We have the opportunity to build a horizontal software — solve this problem in many companies in many industries, and sell them a license to use it. We have the opportunity to build a vertical software — create a software product that works for all of the parties in a given ecosystem (ie data center hardware). The first one is valuable in the short term but not scalable (restricted by our own time and capacity). The second could scale quickly but isn’t long term defensible (labs and/or home built software and/or incumbents will try to eat you and constantly be encroaching). The third will be harder to scale but if you find a niche harder/less likely to replace, but not really because you still find yourself in a similar situation to the horizontal version. What is really valuable and sticky? Owning the communication/the seams because of the network effects that come from having everyone using your thing and your thing being the thing that allows things to work together.
The big value in this supply chain/RMA space is allowing the companies to autonomously exchange information (because that increases speed). The big risk/impediment is reliability of that autonomous information exchange across boundaries (trust).
And so speed and trust seem to be two important values for us. Maybe we call the company Treed. I could get down w that.
So how do we build this defensible comms layer. How do we build the railroad tracks that allow these things to move, the telco cables that allow the messages to be sent, the email protocol that allows you to make sure the right message is going to the right place, but for agents. And how do we make it stick.
For starters to be very explicit, all of the above models (services/consulting, horizontal SaaS, vertical SaaS) are fair play so long as they ladder up to our broader vision. They’re not only fair play, they’re likely indispensable steps in arriving at our vision. The key is to clearly align on the vision, pin it up on the wall for us to see every day, communicate it clearly and compellingly, and then put one foot in front of the other doing things that others might not see or get but that ladder up to that vision.
So now we’ve got NVIDIA cooking. The goal there is for them to implement on their platform, seed us, introduce to clients, etc. NVIDIA is a singularly valuable relationship, not just because they’re a big sexy company, but also 1) because they sit at this critical node in the ecosystem, and 2) because they have a track record of seeding important seminal foundational critical companies in the ecosystem. That’s why it’s critical for bliss and I to talk more about our end vision: is our vision to become the foundation of the ai hardware services ecosystem, or is it to become the medium of trusted exchange between agents. Those are different visions. They walk the same path, but they are different visions and they require different hires along the way.
So we need to push on NVIDIA. We cannot let that go. We need to do everything possible to become one of their darling companies — they fund us, they incubate us, etc. NVIDIA is a singularly valuable relationship towards this vision.
But we’re doing everything we can there. And we will (and must) continue doing everything possible there. So what else can/should we do?
The tricky thing here is how much to plow ahead on the NVIDIA pipeline (ie target the companies to which we want them to introduce us) vs how much to wait until NVIDIA is ready and then go do that.
Ugh. This is where it gets hard. I don’t know. I do think we want to continue having conversations with parties along that NVIDIA chain. My default is totally to think about just attacking the big dogs head on. We got NVIDIA cooking now let’s go get AWS and Foxconn and integrate the whole chain. Yes, that is the right play. A lot easier if NVIDIA makes those intros. But maybe we attack on our own anyway. More on that later — I don’t have clarity.
What if we push ourselves to think in an anti-Dustin default kind of way. We think about how to start small. I mean, still begs the question of whether we start in data centers or not. If we do, who are the small players in the dc hardware space? And are they positioned for growth or are they positioned to die and get eaten? This is another good question for us to consider. Who are the small players in the data center space who are not likely to get immediately squashed.
And do we look more horizontally? If so, what verticals? I mean, isn’t the right answer to just take a half step, to find your way forward, to not allow perfect to be the enemy of progress? Nothing succeeds like success. Nothing will get us further than moving toward. So I feel like we pick up customers where we can find them, including in the medical space. We just don’t know where this will lead.
Anyway, I’m stuck. I need to mediate more on this last point. For our next discussion.