Are we already in the Singularity? | E24

Summary

This Week in AI – Episode 24: Are We Already in the Singularity?

Date: July 30, 2026
Host: Alex Wilhelm (guest hosting for Jason Calacanis)
Guests:

• Grant Lee, CEO of Gamma (AI-powered presentations/websites)
• Philip Johnson, CEO of Starcloud (orbital data centers)

---

Episode Overview

This episode dives into the rapid evolution of the AI industry, exploring where value is accruing in the “AI stack”—from chips and hardware through open-vs-closed models and up to the application layer. The panel reflects on open weights, the economics of infrastructure and compute, shifting regulatory environments, burnout in the sector, and—prompted by Sam Altman—whether we are living in the singularity. The discussion is grounded in real-world founder experiences: scaling, market competition, and the future of their companies.

---

Key Topics and Discussion Points

1. Open Weights Debate and Ecosystem Shifts

[00:22 – 04:48]

• Explosion of support for the open weight model pledge after Nvidia and Jensen Huang joined, with over 100 companies signing on.
• Grant Lee (Gamma): No direct approach to sign the pledge, but supports openness for innovation. “Everybody in the industry…is very big proponent of keeping things open and making sure open has a chance at…continuing to play a big role.”
• Philip Johnson (Starcloud): Open models align with Starcloud’s interests—low cost compute, more demand. Recognizes some virtue signaling but says, “Nobody wants to be on the wrong side of Nvidia…he who has the gold has the rules.”
• Both foresee a world where open and closed models exist simultaneously: “You're going to have both.” (Philip, 04:42)
• Application-layer companies like Gamma benefit from both increased competition and customer desire for optionality.

2. Application vs. Infrastructure: Shifting Preferences

[05:06 – 08:11]

• Enterprises are moving toward AI sovereignty, wanting to build their own stacks, causing model inference preferences to become more fragmented and industry/geography-specific (US vs EU etc.).
• Gamma offers both open and closed models, aims for model-agnosticity and flexibility: “Our own infrastructure is…predicated on experimentation and orchestration.” (Grant, 07:12)
• End users care less about model details—it's results that matter.

3. Starcloud’s Journey: Space as the Next Compute Frontier

[08:11 – 13:34]

• Starcloud sent their first GPU (H100) into orbit, running Nano-GPT on Shakespeare and inference on satellite imagery.
• Philip recounts their humble beginnings, including a lawsuit-induced name change (Lumen Orbit → Starcloud): “We got sued by Lumen Technologies. That's why.” (Philip, 08:58)
• Demonstrated training and inference benchmarks in orbit; future plans involve much heavier launch vehicles as launch costs drop (thanks to Starship, etc.).
• Government/military contracts let Starcloud “tread water” with current prices, but a real explosion requires further reductions in launch (compute) costs: “If starship is flying anytime before like 2032…we’ll be good.” (Philip, 13:34)

4. Burnout and Pace in AI—Maintaining Joy Under Intense Pressure

[14:11 – 29:43]

• Notable Segment: Lillian from Thinking Machines Labs stepping down due to stress; triggers a discussion on founder/exec burnout.
• Both guests stress the importance of team structure and finding joy in the mission:
  • Grant: “It has absolutely been team effort…the way you construct the team is ideally like a table…each leg can be strong in different ways.” (25:01)
  • Recognition that “this is the mandatory pace” in AI; sustaining it depends on problem passion and team support.
• Visual communication is why Grant’s excited about Gamma: “How do I get someone else to see what I see?” “If no one else is going to do it, we feel an obligation to do it.” (28:54)

5. Moonshot’s Kimi-3: Commercial Terms in ‘Open’ Models

[32:43 – 35:00]

• Moonshot released Kimi-3 model weights, but with a twist: serving it as an inference provider requires payments back to Moonshot.
• Grant: “It all comes back to price to value…If things are getting good, people are going to want their, their cut.”
• Philip: “They still need to find energy and infrastructure, right?…we are more thought of as an energy provider.”

6. The Death (or Not) of Fine-Tuning and the Challenge of Model Agnosticism

[37:10 – 41:32]

• Discussion if ‘fine-tuning’ as competitive moat is dead.
  • Grant: “It's not dead yet.”
  • Philip: “I heard fine tuning was dead. Is it not dead?”
  • Ongoing optionality is required, as market shifts could force reliance on open or closed models to change at any time.
• Cost of fine-tuning has decreased, especially with partners; now a viable enhancement for specific tasks.

7. Stripe Rumored to Acquire OpenRouter—for $10B?!

[44:03 – 48:09]

• Stripe rumored to be acquiring OpenRouter (AI API routing company) for $10B.
  • Grant: Brand and user love are hard to replicate. “On the surface it's like, oh, maybe just throw a few engineers at it. But the reality is it's not always that easy.”
  • Panel surprised at the valuation, questioning defensibility/moats.

8. Economic Trends in Compute and AI Models

[48:09 – 51:46]

• The per-hour cost of GPUs (e.g., H100) is holding up or rising, supporting Starcloud’s economics.
• Newer, more power-dense chips are actually easier for Starcloud’s architecture (“The more power-dense the better almost for us.” – Philip, 50:27).
• StarCloud learning from StarCloud 1—low failure/restart rate, better than expected hardening.

9. International AI Markets: India

[53:04 – 56:48]

• Noting special pricing for Indian market from Cursor, OpenAI.
• Grant: Massive market, lots of free users in India. Gamma is considering regional pricing but wants to maintain product quality; more flexibility as more open models become available.

10. The Cost of Compute Thought Experiment

[56:48 – 57:33]

• If compute was truly unlimited, would Gamma be more or less defensible?
  • Grant: “If we were just to stagnate in terms of, hey, just serve, what would the vast majority of PowerPoint decks look like today then? Yeah, then I think our market isn't very interesting…So you basically can ladder up to what's your next level of ambition and try to build the next thing.”

---

Highlight: Sam Altman Says “We Are in the Singularity”

[57:33 – 66:48]

Clip: Sam Altman

“We are now like in the singularity. Like this is the moment… Now we're like actually in the moment that we used to like talk about at the lunch table in a very not serious way… For the last 10 years ago, this was like a kind of far off dream at best.” – Sam Altman, [58:00]

Reactions

• Philip: 100% agrees: “We are very much now past the point of no return. So I would very much agree that we’re in singularity.” (60:33)
  • Defines singularity as recursive self-improvement and runaway exponential curves in any AI-relevant metric (compute, tokens, etc.).
• Grant: Also agrees. “[The singularity] ends up becoming this sort of spectrum. And if you think we're in this, sort of there is this period of exponential, then I don't think there's any denying that we're definitely in it.” (60:40)
• Alex: Frames singularity as “when do machines create net new intelligence faster than humans?”
• Discussion of “when does it get weird?”:
  • Philip: “When robotics takes off. When…there’s a humanoid carrying somebody’s bag.” (62:19)
  • Grant: When self-driving cars become dominant.
  • Alex: Maybe things won’t feel weird due to human adaptability, even as advances accelerate.
• Philip raises a “p(doom)” (probability of AI doom) perspective, referencing Fermi Paradox and suggesting intelligent life might always be short-lived.
  • “My p doom is like 99.9%.” (64:47)
  • Grant: “I feel like I'm maybe just too optimistic about everything. So I appreciate you guys bringing me back to Earth…” (65:37)

---

Notable Quotes & Moments

• Open Models Debate:
  • “Nobody wants to be on the wrong side of Nvidia…he who has the gold has the rules.” – Philip, [03:29]
  • “You're going to have both [open and closed models].” – Philip, [04:42]
• Engineering in Orbit:
  • “We got sued by Lumen Technologies. That's why.” – Philip, [08:58]
  • “If starship is flying anytime before like 2032…we’ll be good.” – Philip, [13:34]
• Application Layer:
  • “Our own infrastructure is…predicated on experimentation and orchestration…” – Grant, [07:12]
  • “If you feel like the team can balance out the pressure then it makes it possible to build long term.” – Grant, [25:01]
• Fine-Tuning Moats:
  • “I heard fine tuning was dead. Is it not dead?” – Philip, [37:10]
  • “It's not dead. It's not dead yet.” – Grant, [37:14]
• Singularity:
  • “We are very much now past the point of no return." – Philip, [60:33]
  • “It's never going to happen in San Francisco.” – Alex on self-driving, [62:43]
  • “My p doom is like 99.9%.” – Philip, [64:47]
• Cultural Touch:
  • Candid moments about founder stress, adaptability, changing signals in venture rounds, and the dynamic between “building what you love” and surviving in hyper-competitive AI.

---

Timestamps of Key Segments

• Open Weights Debate: [00:22 – 04:48]
• Application Layer Trends: [05:06 – 08:11; 17:54 – 19:49]
• Starcloud’s Mission & Space Compute: [08:11 – 13:34; 48:09 – 51:46]
• Startup Burnout & Culture in AI: [14:11 – 29:43]
• Moonshot’s Kimi-3 Terms: [32:43 – 35:00]
• Fine Tuning and Defensibility: [37:10 – 41:32]
• Stripe & OpenRouter Rumors: [44:03 – 48:09]
• India and Regional Dynamics: [53:04 – 56:48]
• Unlimited Compute Thought Experiment: [56:48 – 57:33]
• The Singularity Debate: [57:33 – 66:48]

---

Conclusion

• Both guests are cautiously (or enthusiastically) optimistic about the prospects for their companies but recognize intense market and personal pressures as the industry races ahead.
• In their view, we are now—by most definitions—living within the beginning of the singularity, as investment, capability, and AI impact enter runaway territory.

Where to find them:

• Gamma Careers: careers.gamma.app
• Starcloud Careers: starcloud.com/careers

---

“If you think we're in this…exponential, then I don't think there's any denying that we're definitely in it. And then the question is what comes next? And I guess we'll have to wait and see.”
– Grant Lee, [60:40]

Transcript

[A]

Hey everybody. Welcome back to this week in AI. My name is Alex Wilhelm. I'm really excited to have you here. Now, where will value accrue in the AI game? I love this question. For a long time it appeared that the foundation model labs would eat the world. Today, chips are taking a huge share of profits and memory is another big thing. Or maybe it's going to be other pieces of silicon or hardware. Will hyperscalers escape with the bag or Neo clouds? Will it be companies operating in the application layer or or forward looking compute players that are betting against the current Earthside data center rush? Today we're going to get into all the big questions and the latest and the news cycle has been insane. Sam Altman thinks we're the singularity. Everyone's mad about open weights and potential regulation. Nvidia keeps spending. Microsoft is taking on the AI labs. K3 has commercial terms, AI routing, burnout and more. So to help me unpack all of this, I brought two of our friends.

[B]

Thanks to our friends at PayPal, the exclusive sponsor for for this Week in AI try the payment and growth platform that's trusted by millions of customers worldwide. PayPal Open start growing. Today at PayPalOpen.com we have grant Lee

[A]

from Gamma, a startup that puts AI to work to create brilliant presentations and websites. Grant, glad to have you back man.

[B]

Thanks for having me. It's great to be back.

[A]

And we also have Philip Johnson speaking of repeat guests. He's from Star Cloud, a startup that wants to build orbital data centers so that we can keep our terrestrial farmland. Philip. Hey man, how you been?

[C]

Hey. Very good, how are you? Thanks for having me.

[A]

I'm. I'm really good and I'm glad I have you guys here because I do want to start with the biggest topic and I want to get your take on this. So everyone is riffing about open weight models, regulation thereof. And I thought that this had died down just a little bit, but as it turns out there was a letter that Nvidia dropped, Jensen joined X which everyone was freaking out about. And There were about 25 names on this letter defending and advocating for open models. Since then it's going to over 100 different companies that have all signed up about this. OpenAI joined and we'll talk about Anthropica in a second. But I'm curious about who got essentially asked to sign and if they wanted to or not. So Grant, I know Gamma is a pretty big name now in the AA game. Over 100 million ARR multibillion dollar valuation. Did anyone reach out and say, hey, do you guys want to sign the, the open weight survivability model pledge?

[B]

We have not been reached out to, but it's definitely a topic we're following closely and you know, it's important when Jensen joins X. So everyone's ob keen to see what, what's going on here.

[C]

Would you, would you sign?

[B]

Yeah, I mean I think everybody in the industry, you know, other than maybe the two, two major players is, is very big proponent of keeping things open and making sure open has a chance at, you know, continuing to play a big role in how we innovate in AI and just technology broadly. What about you, Philip?

[A]

Virtue signaling? I'm trying to figure out like, like everyone showing up suddenly to take part in this. Is this bullying, Philip?

[C]

There's certainly some bullying going on. Nobody wants to be on the wrong side of Nvidia. No, obviously we are never almost. I can imagine Grant is it's in everybody's incentive to want open weight models for us, but we're not the guys pouring billions of dollars into R and D to create the weights. So yeah, I mean it came out yesterday that Nvidia is about to plow $250 billion into OpenAI data center like backing or something like that. So it would be difficult for then OpenAI not to sign a letter.

[A]

There's an old joke that he who has the gold has the rules. Or I guess in this case somehow Nvidia has both all the money and the GPUs. They do seem to have quite a lot of market leverage and I want to get to OpenAI and anthropic in a second but I was thinking about how this applies to your guys companies and Philip, in your case, the more open models there are generally speaking, I think lower prices, therefore more AI demand and therefore in time more demand for orbital AI compute. So to me there's kind of a relatively straight line towards this helping your business.

[C]

Yes, correct. We view ourselves as essentially a provider of low cost energy and infrastructure for data centers. And so the more people either open or, or otherwise that want to produce tokens, the better for our business. So there's basically no world where we would want close models, you know, str structurally. I mean, I, I, I also, you know, believe in it as a concept. But just if I were purely to be looking at it structurally from a star cloud perspective, just to double click

[A]

on that because I don't want anyone watching this to get it twisted because everyone's pretty hot right now about people being on the right side of history. You're saying that closed and open models should exist at the same time as co equal players in the global AI race.

[C]

You're going to have both. I don't think there's going to be a world where you're not going to have both. So yeah, okay.

[A]

And then Grant, in your case, I was just thinking about more open models, more overall AI competition, better AI models. And then because that is the intelligence that you bring to bear inside of Gamma, anything that drives more competition between open weight, Closed Weight Labs, et cetera, is good for your business over a long enough time.

[B]

I mean, it's directly good for our business, but I think indirectly just for our customers. I mean, this is what our customers want too. We're working with many large enterprises where companies are all going through this moment of figuring out whether it's AI sovereignty or the whole notion of renting intelligence versus owning it. People are starting to really want to build their own stacks. And so closed and frontier models will play a role in maybe, you know, subset of tasks or maybe, you know, certain areas where it's really important to be at the frontier. But the majority of enterprise tasks probably don't rely on that just yet, or don't need that just yet. So then, you know, how can you build an AI stack that takes advantage of, of everything else, including open? And then for us, you know, we're model agnostic, so we can sit on top of any of that and be really the visualization layer. And so customers can mold their intelligence stack the way they want it to be. And then we can really work with any of that and really allow them to. What we're passionate about is really around visual communication. How do we empower individuals internally to be able to craft and communicate effectively? All of that should be in a way that is in service of the customers and what they were trying to achieve there.

[A]

No, that makes good sense to me. Now I know that Gamma offers a mix of closed and open models. I ran some data on how many you have of each and I think you're slightly weighted towards. If just going based on your documentation. Yeah, a bit more on the closed model side because I think there's more available there. But I'm curious about on the customer side. You know, where do people put their inference today? Are there a clear kind of like preference or preference shift in terms of what models people want to use and are they moving away from those that are closed?

[B]

I mean, it's highly industry specific and geospecific. You know, I think when you think about enterprise, you know, the trend towards wanting to build more of their own stack. I think we're, you know, early stages of what, figuring what that looks like. Because for a long time it's been hard to kind of stand up your own infrastructure. It's getting way easier to do so. And then even when you look at, you know, Europe versus the US versus any other geo, you know, US obviously is heavily concentrated in the two big players. You look at Europe, it's, it's much more fragmented and there's almost more of a tendency to, almost want to ban some of the US kind of closed models. And so I think again, we're, we're so early in it all, it's, it's going to evolve pretty rapidly. Certain industries are much more protective of their ip, which is like they're, they're feeling the existential need to move much faster into building their own stack. In other industries, it's not as much of a concern. So, yeah, I think we're, you know, we want to just. We don't want to be overly prescriptive. We know at this point being adaptable is going to be incredibly important. We've, you know, kind of our own infrastructure is sort of predicated on this notion of experimentation and orchestration and being able to really, you know, leverage the best model for the right task. And at the end of the day, you're trying to just pass that value onto your end users. Your, most of your end users over time aren't going to care what models are powering most of their tasks. They just want it to work. And as long as we can kind of lean into kind of building best in class, state of the art for visual communication, I think we'll be in a good place to continue to win there.

[A]

All right, so Philip, when you launched your first H100 into space, and I think we can probably play the clip of it going off the spaceship right now, and you guys said you used it for inference or trading. I forget which one it was. What model were you running?

[C]

Yeah, before we get into that, just wanted to hark back to the first episode we did with you where you were incredibly helpful and very supportive before anybody else was. So I just wanted to slip in that we are very grateful to your for your continued support of what we're trying to do. We were in a tiny. It was just me, Addie and Ezra in this, like, garage. Essentially at that point we were called

[A]

Lumen Orbit, I recall. I'm still mad about the name change, by the way. Starcloud's lovely, but I liked Lumen. Orbit is just more out there to me anyways.

[C]

Yeah, I quite liked it too. Yeah, we got sued by Lumen Technologies. That's why. I don't know if you ever.

[A]

Yeah, well, that's no fun at all.

[C]

Yeah, we were running, so we trained the first model in space and we also ran the first version of Gemini and we also ran the first high powered inference on satellite imagery. So with the first. Oh, there we go. Boom, there she goes.

[A]

It looks fake. And so cool.

[C]

It's so cool. So half the time this will deploy in the shadow of the earth. You won't see anything. The other half the time it will deploy not into the silhouette of the Earth. So you also don't get that kind of nice shot. So we got super lucky that we got that shot. If you do anything to this in post production, play the sound that there is with that where it's like, star cloud one separation confirmed.

[A]

Star cloud one separation confirmed.

[C]

Really cool.

[A]

Everyone wants to hear themselves discussed in NASA speak. Like, forget the NASDAQ billboard they have in Times Square. Everyone wants to hear that.

[C]

Yeah, yeah. Literally. It's so funny that I kept reposting that and the SpaceX lady that actually did the shout out, like messaged me on LinkedIn. She was like, that's my voice, by the way. It's like the most perfect voice. Anyway, so, yeah, we put on there the Nano. The nano GPT model from Andrej Karpathy. Do you guys remember when he did that video, like a YouTube video explaining how to train on the complete works of Shakespeare and then predict the next letter? So that's what we did. We put the complete works of Shakespeare on the spacecraft and then we trained it on that. Mainly because with 1H100 you can't do too much, but you can at least demonstrate the time it takes to train nano GPT, which has become a benchmark in itself, like training? Nano GPT has become its own benchmark.

[B]

Yeah.

[A]

Don't people run auto researcher against that as a way to kind of prove that we are seeing the early elements of recursive self learning?

[C]

Yes. Yeah, yeah, yeah. And also, yeah, the faster you can do that, it's kind of a good benchmark for how fast you can do everything else. So if you can train to a certain level of depth in like three minutes versus two minutes versus one minute, you know, then that's what people are trying to do. So yeah, we did that. We then ran Gemma, which is like a cut down version of Gemini in Collaboration with Google DeepMind had a very nice tweet from Demis to Savis about that. And then we also are now doing high powered inference on SAR data, so synthetic aperture radar data, and that's kind of the first useful application is to run inference on satellite imagery, either hyperspectral or high res imagery, and also on other types of sensing data like synthetic aperture radar.

[A]

So things are going apace at ye olde.

[C]

Things are going apace.

[A]

Lumen orbit.

[C]

We've just won four in the last two months. We've run four different government military contracts. Wow. Things are, things are chipping.

[A]

Wow. I just also read that the government's going to put hyperscale data centers on military bases. I guess they're also going to be putting them in space with y'.

[C]

All.

[A]

Can we Talk about the SpaceX IPO just for a hot second? I know this is the speaking AI, speaking wrapping some AI ness, but I was just thinking about your company when they listed and also with the recent starship launched, because going back to that first interview with you, you told me very clearly that look, if Starcloud is going to succeed, we have to get launch costs down. Otherwise there's just not the economics to make this entire thing functional. So I think a lot about compute and we're going to get to the Nvidia thing in a second that more or less to the Gamma side of things. But talk to me about your pace of development of the heavier launch vehicle and how you're feeling about the long term or medium term economics of Starcloud.

[C]

So the business can run with a sort of, you know, $500 million revenue, billion dollars revenue just on this government military use case. So we can sort of tread water for the next, you know, even if it was the next five years with the Falcon 9 launch cost and that would be fine to get to sort of venture scale with, you know, hundreds of billions of dollars of revenue. We do need much lower launch cost, so at least 10x reduction from where we are today. And so that does require something like Starship with a reasonable upstage or maybe stokespace or some of the others. It looks like. Yes. I mean Starship is like on track as far as I can see. They're building these two starship gigafactories designed to produce like three starships per day. So you know, even if you, even if it's a tenth of that, it will be mind blowing. So yeah, if starship is flying anytime before like 2032, like right, then we'll Be good.

[A]

Like I did not realize that we had that much time. I feel like everything in AI operates in like two week sprints, it feels like. So 2032 is five and a half, six years from now. That's an insane timeline. That's.

[C]

I mean, ideally it'll be much sooner than that, but like, you know, they're saying end of 28 could be early 29 before they start having the first commercial payloads. That would be great. The point is like we can survive. You know, we've just done an extension to our series A which we'll announce in a couple of weeks. But so we've got enough cash for, you know, to survive that amount of time. And yeah, it looks like they'll get there at some point. Maybe it will take a little bit longer.

[A]

In the old days of venture, extensions were always kind of looked down upon. Now I think in the AI era they are actually just a way to accelerate companies. It's interesting how the signal of certain things has flipped entirely on its head. Like if you go back far enough, inside rounds were considered to be a bad signal. You couldn't find another lead. But now the biggest signal is, oh, I don't know, Andreessen is leading our Series A, B and C because they want maximum allocation. So many things that used to be common sense have now been flipped on their heads. I guess also like you should put your compute on the ground. No, now we're going to put it in space. You know, a complete inversion of decades of history. Philip.

[C]

Yeah, well, just to clarify, some extension rounds I think are a bad signal. If, for example, if you're doing an extension on a safe with no valuation increase, that's probably not a great signal. Ours was a priced extension of 2x valuation of the previous round. So it, well, I mean we could call it Series A, except there's no board member and it's like basically the same docs as a series A Series B. Sorry. Except there's no new board member. So we decided, and it's like very close. So we decided to call it an extension. But I don't know, I mean at

[A]

that point you just, you're just using terms. I mean seed rounds used to be like 500k, not 500 million.

[B]

Alex, were you at, were you at

[A]

the raise conference, the recent one? No, sadly I have a six month old at home, so I'm pretty landlocked.

[B]

You should have seen Philip. Philip was the. Everybody was just flocking to talk to him. They were. So, you know, everyone else is working on agents and apps and like, you got. You got Philip over here. Just swarms of people.

[C]

This is not true. So. So for the listeners, Grant. Grant and I ended up being on the same table at dinner. And I was like. I was like, oh, my God, that's Grant Lee, the guy from, like, from Gamma. Because me and my girlfriend, me and my fiance now are like, big users of Gamma.

[A]

So I was like, I've got to

[C]

go get a picture and, like, send it to my fiance that I'm sitting on the same table as Grantly. I ended up sitting on the. On a table with Grant at the raised dinner in the Louvre. Like, no. The Chateau de Versailles. Yeah, yeah. And I was like, fuck, is that Grant Lee? Oh, my God. Because we, me and my fiance are like, heavy users or we're big, big fans of Gamma. So I was like, shit, I've got to go take a picture with Grant and send it to my fiance because she's going to be so impressed. So that was how. That was how it went down. It was me wanting to take a picture with Grant more than anything else.

[A]

Well, the problem is if your spouse is not in technology and mine is in medicine versus tech, nothing that I do ever has any kind of impact on her. But when I knew about what's the Dr. AI thing, open evidence or whatever. Yeah. She was like, oh, yeah, that's great. And I was like, I'm finally relevant to the conversation.

[B]

One tool that broke through. Yeah.

[A]

But just to be clear, you guys were at the. The Versailles Palace. Yes.

[C]

Yes.

[A]

That's a huge upgrade. I once went to a.

[B]

Underground.

[A]

Yeah, I once went to a palace in London is one of the lesser ones. And I just. Walking through it, I just had this huge feeling of, like, institutional decay over the centuries. Like, it just felt like I was in this old building and it's so weird to have, like, tech people, like, next to swords. It just didn't feel right in a way. Was. Was Versailles fun?

[B]

I mean, I feel like I had fun. Yeah, Same here. I got to give credit to the. Yeah, the conference organizers. I felt like they did a great job. It's really hard to pull off something at that scale. And it sounds like they actually want to double it next year, so we'll see. But I feel like this year was actually the right sort of balance because a lot of great attendees and the concentration just made it so that the meetings were great. You can have back to back, and it all was in one place. So we'll see what they do next year. But Got to give them a shout out for this year they want done.

[A]

Yeah, getting people together is always such a good use of time. So if Philip was the, the celebrity, then. Grant, I'm curious what people wanted to talk to you about at this event because I don't think we talk about the app layer enough when it comes to the AI conversation. It almost feels like it's something that's set to the side compared to talking about, I think what we now call primitives, you know, compute data and so forth. And so were you, were you less in demand because people don't care as much about the app layer or is there something else going on there? Because I think this is actually we're joking, but it's an interesting signal from the market.

[B]

I mean, I think the reality is that there's so many, you know, the attendees there were just off the charts. So everybody there was, you know, impressive backgrounds, all, all working on really impressive businesses. So it was just a great company to be in. I think everyone, you know, the app layer, certainly right now everyone's trying to figure out, you know, what is the sort of long term approach to, you know, open versus closed. How much of it do you, you know, change what is working well today. Oftentimes, you know, if you're leaning on one of the frontier models, you've just been all in for, for so long. So how much of that mix really needs to change? Or do you just keep riding that for, for as long as needed and then, you know, worry about the, you know, downstream, you know, stuff later on? I think all conversations kind of predicate around that for me. And, and then when you think about like the enterprise buyers, it just still feels so early. You know, obviously on the prosumer side, that's where most of the tailwinds has been, you know, for, for most of these app and agent level companies for the past year, two years. And now we're starting to see real enterprise appetite. And so everyone's trying to figure out, you know, is this a land grab for their specific segment and how much do they really need to lean in and how do they compete? So these are all the sort of ongoing conversations. And I would say yeah, come this summer, really feels like things have really escalated. Like the demand is there, the competition is definitely increasing. So everyone's trying to figure out the right swim lane.

[A]

Yeah, and this is why I think your series B, or we could call it your series A extension if you want to go with Philip nomenclature here. You guys announced that you hit 100 million ARR. And this, by the way, this is last November data. So it's definitely behind. Grant has not blessed us with updated metrics that I know of. You were at 100 million RR, you raised 68 million $2.1 billion valuation. Hold on to that number for a second. But also you talked about like profitability and having fun and so forth. And I thought that was a really interesting kind of approach to building a company. But when you're in that space, you're all these other founders. How serious was it versus how enjoyable was it? Because I think that people are grinding real hard and losing some of the spark. And I was really encouraged just prepping for this, hearing about how you had a very different approach to building the company.

[B]

Yeah, I mean, I think, you know, for us we've tried to keep the team pretty lean and that in itself just, you know, changes the dynamic. But every company needs to at this moment in time, like real recognize like if you're going after global opportunity, you know, you can't keep headcount, you know, kind of capped forever. You definitely need to grow. We're going through that today. If you're going after Enterprise, it requires a different go to market motion. So you have to scale the team. So I think you always want to balance like hey, you know, hard work and fun. Ideally those two can coexist in an organization. And if you're building, you know, strong culture, then you can, you know, hold on to whatever version of that makes sense for you. And so yeah, we're going through it all. I think competition has definitely increased since, you know, since the Series B. And so the question is like, how do we continue to navigate which markets matter for us and as we scale up the team, how do we preserve what we feel like is special about building Gamma?

[A]

So I'd be curious about maintaining the sense of joy and fun and going to sell it to the enterprise because I think if you're more consumer prosumer, which is by the way how I thought about Gamma originally. Glad to see you're going up market. But to me I'm curious if you can maintain and hold on to that as you go and approach these more drawn out contract negotiations and SLAs and enterprise sales because that just seems to the fun out of business.

[B]

I think it's a little bit easier for a category like ours because we oftentimes we go into an organization even though, you know, now coming tops down, we already have users there and so they're championing the product and that part of it is, you know, way easier having a conversation where you're. You kind of have a warm reception versus going in cold is different. And when we have a champion that is like, really fighting for us to be, you know, you know, wall to wall, then that changes the dynamic and we can go in and really offer kind of in contrast to like, a lot of AI platforms that don't give much, you know, human love or touch or like you're. You're not like, onboarding them or like actually doing any sort of level of tr. We go in and it feels like white glove service and they get so much more. So our team gets a lot of joy out of that. So that part of it's still fun. Obviously, as you get further and further along, more competition, more sort of, you have to do whatever it takes. Of course, that might be a little bit more draining, but we're still at the early stages of that, so we'll see how long that can kind of carry us forward.

[A]

Philip, going back to your. If Starship is launching frequently by 2032, everything's going to be fine. You obviously have pretty long time horizons because you're working with space. So it's a little bit different than just shipping software. How do you balance working as fast as possible, taking advantage of the moment, capitalizing on AI and also ensuring that your team kind of doesn't burn out and blow away. And I promise I'm going somewhere with this, but I'm really curious about how you guys deal with this. On the hardware side maybe.

[C]

Yeah, I don't want to give the wrong impression with 2032. That is not the. We have three launches coming up in the next year. So, like, the team is working like, flat out right now. But yeah, we usually give the team a bit like of a, you know, like a week or two of less intense work immediately after shipping the spacecraft. So we're going to ship StarCloud 2 in a couple of weeks and we'll probably. Then, you know, people can go on vacation and all that stuff.

[A]

Okay. Yeah, because the context here. And we're going to pull up a screenshot here. Lillian, a co founder over at Thinking Machines Lab, which has come out and really put out some really awesome products in the last six months. She's stepping down from her. Her perch across, you know, atop one of these, I mean, frankly, top, you know, n of, I don't know, five companies out there. And she just talks about stress. Yeah. And how, you know, she just didn't feel like she couldn't give 100%. And she's actually leaving the company and her co workers on X really, you know, an outpouring of affection and love for her. So clearly a very amicable exit. But this shocked me and it made me a little bit worried about how people are doing that are building these companies because I sit in my chair and I go, oh Gamma, you know, 100 million ARR under such a short time period, Hell yeah. What an amazing accomplishment. What, what a business building experiment. And then I'm just seeing this, I worry a little bit about the people and if we're, if we're all trying to do maybe 10% too much and not taking care of ourselves or, or one another, if that makes sense. Grant, how do you take care of the group to ensure that they have the right balance here you don't have a launch so you can't give people time off afterwards. But what's your approach to that?

[B]

Yeah, yeah, I mean reading that you reflect building a company is all consuming and I personally there's no way I could do it without my two co founders. So for us it has absolutely been team effort and I think the way you construct the team is ideally like a table or something sturdy like each leg can be strong in different ways and for us very complementary skill sets. We worked together for over five plus years at another startup called Optimizely and then now been building together for over a decade and there's going to be ups and downs. I think if you feel like the team can balance out the pressure then it makes it possible to build long term. But I have seen you know, whether it's like solo founders or founders that just never didn't have like a surround themselves with a complimentary team that just puts the sole focus on just like somebody that has to deal with all the different fires are constantly burning and I can imagine that just leading to burnout in this environment, especially when there is no certainty, no sort of ability to long term forecast. It's all sort of month, quarter over quarter like adapting at the very least and, and hopefully you have a team that can actually, you know, do that with you and you're not, you're not on your own.

[A]

Yeah, I, I saw some notes from. Who's the guy who does the pragmatic engineer, Gregory, I think his name is. He was talking about how like easing a lot of senior talented companies on the software side essentially either like a step back like CTOs, VP of Eng, either like not take roles or leave them very quickly for a couple of reasons. Bernard is one of them. And then another reason was no one wants to work at a company today that isn't at the cutting edge because it's kind of resume poison. And so people are just taking a break as things evolve. And it just doesn't seem like a healthy way for the industry to capitalize on its existing talent, to move as fast as we can. And so it feels like the short term pressures of the AI moment are almost driving dislocations in labor that, I mean, not everything is USA versus China, but I don't think it helps our national competitiveness grant to, to not have our, our, our best and brightest in the saddle. And so I just, I just wonder what we can do to still win but not burn out so many bright lights along the way. And I wonder if we're trying to do too much too fast. Like, do you think we could go 10% slower and still win, or is this the, the mandatory pace?

[B]

Unfortunately, I think this is the mandatory pace, but how you maintain the pace is, is, is up to you. And so I think many people need to figure out, like, what does that mean to like, go from, you know, progress being measured in sort of years to progress being measured in weeks or months? And how do you make sure you can sustain that for a long time? And hopefully, like, if people are building companies, you know, I think, you know, in the US and what I'm seeing a lot of startups, you know, they have the opportunity to build something that they're hopefully actually passionate about. So going back to like, is this fun? For me, if you're in a category where the problems you're working on are things that you would have been thinking about anyways, or you wake up wanting to think about that alleviates at least part of the pressure of having to do that for a long period of time. The worst case scenario is you get boxed into something that had initial traction, but you actually had no love for the space or care anything about the future or progress in that realm. And now you're handcuffed to it for a long time because you raised money and now you're stuck. And so that's my advice, like all early stage founders is like, yeah, just be prepared. It's going to be a tough and long road and you can make it less draining if you actually pick a problem you actually care about.

[A]

Why is visual intelligence the thing that you want to spend your blood, sweat and tears working on right now?

[B]

I mean, for us, we think about how human progress is predicated or depends on communication. It's like, how do I get someone else to see what I see? This is a moment of extreme opportunity to innovate and actually drive human progress. And it's not everybody that's going to necessarily need a tool like ours. But for those that need to be able to visually communicate, whether it's to their peer, colleagues, bosses, to customers, we want to be the best solution out there and we want to be able to push the medium forward. I think many people get stuck in wanting to just live in 16 by 9 slides. We know that's not what things are going to look like in 10 years. Right. And who is going to be the one to invent what that visual communication, what that language could look like? You know, if no one else is going to do it, we feel an obligation to do it because we really care about that problem space.

[A]

How long until when I log into Gamma, it's not just websites and slides per se, but like basically anything that I can create, anything that I want to create that has a communication element in it. Because I'm thinking animations, art, long form videos, like I think there's so much in the world of visual that you can expand to because I know you guys are kind of pigeonholed in some people's minds as the AI presentation company, but you're already doing more. And so I'm curious, how wide does that aperture get and still kind of sit neatly underneath the umbrella of visual intelligence?

[B]

It's going to get very wide for us just in the next couple of months. Like we're going to be releasing a ton, I think, where the real, you know, what we're excited about is obviously what goes beyond what exists today. Like, you know, there's already formats like PDFs and PowerPoints that have existed for a long time. What, what goes beyond that or what comes next? That part of it is yet to be written and there's a lot of things we dog food internally are tinkering with. We're excited to gradually roll some of those things out to see how the market might react. But, you know, that's part of the fun. It's a, it's a, you know, that game is something we want to play for a long time.

[A]

Yeah. So on this point, Philip, about, you know, pursuing joy, doing things that you love, I kind of think about your company in a different light because to me it's a very binary bet on the world, terrestrial world remain compute constrained and everything that I can see up through Alphabet's earnings, by the time this comes out, Microsoft Will have dropped. We are recording just before that. So I don't have that information top of hand, but it does seem that everyone's still compute constrained and that is a really lovely tailwind or driver for you. Is that enough or are there other things about building basically space based compute that bring kind of a tear of joy to your eyes?

[C]

So one of the. Yeah, as I said, we can sort of survive indefinitely just on serving, you know, providing inference and other cloud services to other spacecraft. And you know, I think that's a very good market to serve. You know, like there's lots of very good use cases for that, you know, search and rescue being one of them. If you have satellite imagery on, you're trying to locate a vessel instead of having to wait three days to get that data back because you have to wait for ground station passes, we can run high powered inference on that data on orbit and draw insights from that way quicker. And same with Wildfire detection, same with many different use cases like this. So, yeah, it's definitely a use case that I would get passionate about. I think it would be much harder to get venture capitalists as interested if it was just going to be that use case. But that's not really my problem.

[A]

Not really your problem in that you're not seeing a shortage of interest in venture capitalists?

[C]

No, no, I didn't mean it like that. I'm all meant like, if VCs decide they're not interested in this because suddenly the compute isn't constrained, that's fine. We'll just build this business for providing cloud and edge services for other spacecraft and I'll be happy and have fun doing that regardless.

[A]

Well, that's a great perspective to have going into your eventual Series B. I don't need you guys, but you can come on board if you want. All right, let's put aside the philosophy for a minute and talk about some stuff that matters a bit more concretely. So moonshot dropped Kimik 3. And when they did, they said, hey, we're going to put the weights out on July 27th. And they did. With a twist. The twist being that with this new model and its new weights and its new license, if you want to serve it as an inference provider, you have to pay something back to Moonshot. Those terms are private, but it does seem, grant, that we're seeing companies building open models that we might have just expected to be able to use for free are no longer offering them up in the same way. So from the Gamma perspective, being model agnostic, how does changing Terms in the open weight game impact what you want to offer to your customers.

[B]

Yeah, I mean, it all comes back to like price to value. And you know, we want to have obviously maximum optionality. We want the best models out there to be able to choose from and then be able to say in which cases do, does the price still align with the value we can actually pass on to, to our, you know, to our users? And so, yeah, I mean, ultimately, I don't think any of this is surprising. I think it was a matter of time. If things are getting good, people are going to want their, their cut. And if, and if, and it is good, they deserve their, their cut. And so then the market will, you know, to kind of dictate how does the overall mix work for, you know, across the entire sort of, you know, set of use cases. And for us being like one of the players in the space, I think we'll have to make our own decision in terms of what mix actually makes sense.

[A]

Okay, so Philip, you know, we were talking earlier about why StarCloud likes open models and how it's going to be overall good for your business. I was just thinking about you though, because if you were, let's say, an inference provider from space and you were serving models to trust your customers in the way we think of them now. So in that context, if you're competing with Moonshot and have to pay them a cut of your revenue, that puts you at a pretty material price disadvantage. And so I was curious if this new approach makes Star Cloud as a token factory less appealing economically.

[C]

The way I view that relationship with Moonshot is they would pay us for energy and infrastructure and then they can serve or anybody else can then pay for. They still need to find energy and infrastructure, right? Yeah. So maybe if, let's say for example, maybe on AWS or one of the clouds, you can opt in to use Kimi and pay that fee that will get either get passed through to the customer or to the cloud provider. But the cost of energy and infrastructure is kind of commoditized or, you know, it's comparable across different places that you can get that. So if we can provide that cheaper, we should still be able to charge the same price that it would cost terrestrially, so.

[A]

Well, no, I think the terrestrial versus space based will sort itself out. I was more thinking about you guys versus Moonshot, but it sounds like my mistake there was thinking about you guys as a inference provider versus a NEO cloud. Essentially.

[C]

Yes.

[A]

Okay, so you guys are going to just be the equivalent of bare Metal in space?

[C]

Yes. Maybe not even bare metal, maybe just energy and infrastructure. So like we, we work with, for example, the relationship we have with crew servers. The agreement we have there is we have a box and that box has power, cooling and connectivity. And you can do whatever you, you know, we'll work with you on putting whatever chip architecture you want in there. You can sell to whatever customers you like for whatever price you like. And we are more. More thought of in that case as like an energy provider, whereas like a cruise, whereas like a, a core weave are buying the chips and then other cloud providers are running on those chips.

[A]

Huh. Grant, for, for your inference needs, how do you approach apportioning that between Hyperscalers, Foundation Labs and NeoClouds?

[B]

Right now it's a lot on Frontier Models Labs and then working with partners like Base 10 to not only serve but also train fine tune models that we can then figure out when and how to deploy. So again, this is all moving target in terms of what that mix actually looks like and what the dependence looks like on any one of these, you know, providers and hard to say what things might well even look like a year from now.

[A]

Did you mention base 10 there? Oh, Philip, please.

[C]

I. I heard fine tuning was dead. Is. Is it not dead?

[B]

It's. It's not dead. It's not dead yet. No.

[A]

We.

[B]

Yeah.

[A]

Philip, explain why fine tuning might be dead. This came up on a venture capital roundtable I did the other week. But I'm curious why you think that fine tuning may be going the way of the dodo for now.

[C]

Well, I, I think people used to think fine tuning was going to be their moat and it was like a very nice story for VCs. We have all this proprietary data that we can use, but it turns out the Frontier Labs are way better at like legal tech or, you know, like, for example, with Lagora. I had Max speaking at YC Startup School on the weekend and he was saying everyone thought that moat was going to be. We have all this legal data and that's going to help us fine tune models. Turns out Anthropic's way better at writing legal models than we ever would be.

[A]

So it's kind of a humbling moment. The thing that I heard from VCs was that. But also that because new models are coming out so quickly, by the time you finish your, I don't know, post training or fine tuning run, you only have this short window of time before you essentially have to start over again. And that companies just can't afford to give base 10 or fireworks that much of their revenue. And so I'm curious, Grant, you're nodding your head here. Why are these complaints wrong and why is fine tune still in?

[B]

It's not that. Yeah, I agree. The sort of sentiment shifted from like this is going to be our silver bullet or like long term defensibility. I don't think that's necessarily true. I think when it comes to very specialized tasks, you can imagine like within visual communication, like all the things that need to be done right, to synthesize information, to pick the right type of visual, to construct the visual, there's a lot of things we can do to start fine tuning models to do specific tasks for really, really well and faster and cheaper.

[C]

And is that not just better prompting, like, is that. Not just.

[B]

That alone gets you decently far. But then of course now if we're using open models, then we can, you know, we have much more ability to start experimenting quite a bit and then, then you can actually change the cost equation. And so right now, you know, we can start using a lot of open models to get almost the same performance for a variety of tasks. And maybe over time the ability to just know how to actually fine tune just specifically what you need, if that becomes like a core competency, then anytime something, you know, one of the new open models has like step function changes, you can pull that in really fast. And you're, you're, you're kind of predicting that 612 months, maybe a year down the road where, you know, maybe closed models in terms of where they're innovating and what's useful for your domain maybe starts diminishing, you'll have the ability to go back and say, okay, we can actually fine tune any of the open models that eventually catch up at a significantly lower rate and serve that in a way that makes sense for your, for your business.

[A]

That implies though that the closed and open models of the world will maintain a relatively large price delta. Do you think that's always going to be the case?

[B]

Probably not. And I think this is where you can't, I think everyone's trying to make sure that they have optionality. Nobody knows what you know, the market's so dynamic. Nobody knows like of course, when anthropic and OpenAI, if they come under much more, you know, if there's competition in their market and they have to do whatever it takes to, you know, protect our home turf, of course pricing is going to be one of the things they explore first or unbundling some of the you know, you know, their own model offerings. And so I think that's something that may happen, but the timing of which, who knows? And, and because of that, you know, everybody needs to have optionality. So how do you make sure that no single vendor, you know, is, is dominating your, your strategy or the way you can actually serve your customers? So you have to basically start explaining exploring today and making sure that, that again, if you think that there's some world where that needs to be a core competency, you cannot wait a year or two years from now to start investing in it. You have to start investing in it today.

[A]

Yeah. How much does it cost to get started in that? So when you guys were just getting into the fine tuning game, what was your kind of expectation for cost versus benefit?

[B]

I mean, this is where it used to be really hard to stand certain things up. But with the right partners, you know, I was bunching base 10, you know, obviously you can start doing things, kind of remove a lot of the upfront costs if you can partner closely with them. And, you know, you're not talking about, you know, tens of millions anymore, you're talking about much less than that. And so, you know, I think that's where you're trying to see, you know, at which point, like, does that partnership really get you far enough ahead so that you can start building more and more muscle internally?

[A]

My concern is that we're going to end up using third party companies who are brilliant by the way, and do a great job at operating this stuff. Tinker from Thinking Machines Labs is another product in this domain. Shout out great, great people. But we just spent a lot of time arguing as an industry about whether we should use open weight models or closed source models. And people are moving kind of more towards open weight and bringing things internally. And then now we're sitting here saying you should go to this other company that does special magic sauce in their black box for you, so that way you can do more things internally yourself. It just seems like a regression to the same point. And I wonder if optionality and internal capacity or capability grant are going to become like the real currency in a few years. Because look back six months, Google had Gemini three. We were all thinking that OpenAI was cooked. Then anthropic took over and somehow in a quarter and a half's time, now everyone hates anthropic. Like, I mean the pace of change is so great that I, I wonder if flexibility will become the new spine, if that makes sense.

[B]

My prediction is you're just going to see many different types of businesses get formed. I think a lot of SaaS companies end up looking very similar over time. The way they constructed the go to market teams, engineering teams, you look at the ratios of how much of the organization's R&D versus not. I think it's going to be a much broader spectrum of companies going forward. Some are going to make the decision to lean heavily into Frontier and go all in on that. And that's okay because the rest of their cost structure makes sense and they're serving in industries where they don't care necessarily and are maybe not as price sensitive. Whereas many other niche players or the sort of development shops are just much smaller and want to run lean and agile and can spin together 10 different apps or agents at the same time. They might have a very different tech stack and their willingness to depend on different partners is probably much higher. So I think it's ultimately good because I think then it's up to the founders and the company to kind of mold or shape the business that they actually want to build. And going back to the burnout problem, it's, it's the worst. If you're building a company that doesn't actually even resemble the type of company you're excited about, you walk into the door, you know, to the door every day. It's like, what am I even staring at? I think that, I think a lot of SaaS founders actually got trapped into that, you know, reality because they were forced to build a certain way. I think optionality and having more tools and partners at your disposal means you can kind of shape it in a much more, you know, meaningful way than you ever could in the past.

[A]

Yeah, I mean if you think back to the era of big SaaS, I don't think people were that excited about building, you know, the next E signature tool, but people did, they raised money, they went to market, they ground super hard, they ate all the glass to become number two to DocuSign, a company that we all keep expecting to be, you know, obviated by Vibe Coded solutions. I was trying to work our work our way towards a news item that I'm still chewing on quite a lot, which is that Stripe is considering buying open router for $10 billion. And we talk a lot about moat in the AI era and what it means is it Velocity, is it, you know, especially fine tuned models, what is it? And OpenRider to me is one of those companies that like, I freaking love, I love their data. I'm actually an OpenRunner customer myself for my own openclaw setup. Like really this could not say enough nice things about the business but I've always been a little bit worried about defensibility and yet here comes Stripe, potentially allegedly reportedly going to invest up to $10 billion in it. So I just feel like I don't understand where the moats are and where the value is accruing. And that's why I was kind of glad to have you two on because different sides of the industry, the application layer versus future infra. But I don't know which one of you is making the better long term bet. So Philip, where does value accrue? And if you were spending your own money, what's the max valuation you would buy into open router for?

[C]

To be honest, I've been a bit out of the open router discussion so like I have no idea what their revenue is or like how long they've been going or how many people they have. I know it's a good tool, like people talk, say nice things about it but like I'm, I'm a bit.

[A]

Last validation was 1.3 billion. I think so.

[C]

Really? And stripe's 3x jump, well allegedly thinking

[A]

about paying 10 10.

[C]

So cool give for the founders. Why can't Stripe build that themselves?

[A]

That's kind of my question because it's not, I think technically impossible. There are other startups doing this, but yet they seem to have quite a lot of, I mean brand essentially. I think about them, I use their data every day. But I guess I didn't think that Brand was going to have so much weight in the era of increasingly intelligent models and lower cost intelligence in general.

[C]

These deals sometimes behind closed doors can have the weirdest rationales. Like it could be that Patrick Collison has been a daily user since day one and he just is like I love these guys. Like I want the team and like, you know, it's really hard to know what's going on behind these deals.

[A]

Sometimes producer producer Lon weighs in that the the ARR of Open rider was about 50 million in April and he reminds me that Stripe already handles their payments. So when I first saw this news guys, I thought oh, they want to buy essentially a chunk of the AI transactions and you know, make, make their own margins on that. But they were already doing that that and we already know their growth numbers because you can see it in their usage charts. So like nothing here is particularly hidden. So they're not buying something that's blowing up behind our backs. It just, I don't know, maybe brand matters more Grant and therefore all the work you've done to make Gamma synonymous with presentations and slides and now websites is going to yield a real defensible moat.

[B]

I mean I think Stripe realizes that it's, yeah, they can build a lot of things themselves but it's, it's, it's, it's hard, it's really hard to build something that becomes, can become a beloved product or something that is so good, you know, people are telling, you know, colleagues or friends about and you know, they, they, I think they bought like Metronome which is in the sort of, you know, usage based billing platform. And again, yeah, maybe they could have built that themselves but I think they saw a team that was able to do something special in the space and the customers are happy. And I think OpenRE is probably another example is like one kind of, you know, accelerates maybe the roadmap and two, like it just, it oftentimes is just that much harder. Like on the, on the surface it's like, oh, maybe just throw a few engineers at it. But the reality is it's not always that easy. And so they probably have spent a good amount of time with that team or have been just thinking about it and can this just be a nice little pickup for them at a time maybe also where the founders are like, okay, yeah, why not join a team that they also respect and want to be on the longer term journey with?

[A]

Well, I mean everyone loves Stripe and shout out the Carlson brothers and all of that but, but I don't know if I'd want to seed my, my hot startup, you know, being in charge of it for being kind of like what a mid level VP at Stripe. Like it just seems like a pretty substantial downgrade and if they're already doing the payments, I don't know. Okay, so let's do a couple things that are specific to Yalls companies. First of all, Philip, on the starcloud front, I was reading about Google's latest earnings report and it seems that depreciation in gpos is picking up a little bit, but not super materially, not super crazily. And so I know you're talking more about power in space up in orbit, but I do think about you guys still having the first GPU in space. So has the economic longevity of older GPU generations made the viability of space based compute more viable?

[C]

Yes.

[A]

How much?

[C]

A lot. I mean if you stop using the chips at that five years versus seven years, that's a big difference. Especially since now revenue per hour of H100 is higher now than it was two years ago when they came out. So what, two and, two and a half years ago? Probably so. I mean that is a very weird dynamic that is quite specific. It's like an inverted yield curve that's like. It's very supply and demand driven. It's very unlikely to remain like this for a very long time, I don't think. But yeah, I mean if you could, if that was going to continue, it would very, very significantly push the. Push things in favor of doing things in space.

[B]

Yeah.

[A]

I've been watching the GPU per hourly cost charts and I've just been absolutely shocked that instead of seeing them go down, they all seem to be trending up a little bit even after all this time and even after successive generations of getting better and better gear from Nvidia. Which is the other thing I wanted to ask you about because it seems that when I hear about new compute tools, they're getting bigger, they're getting harder to cool, they're getting more complicated and therefore, I presume, harder to shield, harder to harden for an in orbit environment. So as the compute market itself has evolved, has that changed how you think about designing your future satellite clusters and farms?

[D]

Really?

[C]

Yes. So actually like the more power dense the better almost for us. So the like, for example, if we can just launch one NVL72 rack, which is essentially what we're planning to do for StartCloud 3. So that's like on the order of 150kW depending on specs. If that, you know, goes up as one node, like one StarCloud 3 satellite, that's good. Like it's a good architecture for us. It doesn't make actually sort of the more bulky and power dens they get, the easier things are on shielding. And also like we, because we have this directed chip liquid cooling architecture going out to these large radiators, like it's not a problem for us having very power dense chips.

[A]

That's very interesting. So these new chips make things more viable for starcloud, not less overall. Has there been anything that you've learned in the last couple of quarters that makes your project seem more difficult? Because there does seem to be quite a lot of tailwinds, launch costs falling, you know, more power, concentrated GPUs. Is there anything on the other side of that coin that's making you more worried? Micrometeorites, radiation cooling?

[C]

Do you mean findings from Star Cloud One or from StarCloud One

[A]

behind you? I can see the engineering process going on. So I'M just curious if there's any less positive news for the future of In Orbit Compute.

[C]

No, I mean, the more we get into it, the more so, you know, we start with a whole bunch of engineering challenges and we're whittling them down as we go. So if anything, I'm a lot more positive now even than when I started. Especially since StarCloud 1 is still performing miraculously well. Like, you know, we're still doing demos for the military on starcloud one now and it's like we've not had a single. No, I tell we've had one restart failure on StarCloud One so far, but we were expecting that to happen like every two weeks.

[A]

Restart failure. Tell me what that is.

[C]

Like where for example, if you have a heavy iron that hits the power supply or hits something else, you might need to restart the chips.

[A]

So radiation in space essentially forcing you to reboot the chip itself. Yeah, there's a famous example of someone played speedrunning a video game and then there was a bug that no one could ever replicate. And the idea is that probably it was a spare bit of space radiation that caused a bit flip inside the computer that was running the game that led to this particular issue. So I'm familiar with the idea in the context of the gaming world. I just didn't think that it would apply as much to what you guys do because of hardening and shielding. I thought that would preclude all those issues.

[C]

Well, we're designing it to preclude all those issues. Yeah, we'll have much more shielding on StarCloud 2 than we did on StarCloud 1, but yeah, it's a process.

[A]

Well, I'm really glad there's not a lot of sand in the gears. I mean, you're doing something that's incredibly hard and to have things get easier and better looking as we go along is exactly what we want to see if we're going to get our compute cost down to essentially de minimis. Grant, a couple things that I wanted to run by you. Cursor just announced special pricing for the Indian market and we already know that OpenAI has built India specific subscription tiers. I think it's ChatGPT Go was the name of that. And I'm curious about what you guys see from the kind of consumer prosumer side of your business. I know you're going more towards the enterprise side, but how important is the Indian market and how much demand are we seeing for AI services and products from that particular country?

[B]

Yeah, it's a massive market, we see a lot of demand. I think we haven't been able to evolve our pricing and packing as much as we'd like to. So I think being able to offer different plan types or tiers specifically for different regions that we can still deliver in a way that is sort of protect some of the margins that we have as a business. I think that these are things we want to be able to do. I think going back to like whether or not, you know, being able to lean more into open models allows us to have more flexibility. I think it's something we're actively exploring and, and I think it's going back to like, if you, if you're, if your ambition is to be, you know, global business that serves, you know, countries and regions around the world, then having the optionality and figuring out how you can actually construct pricing and packaging that works at that scale, you know, that that's what we have to still figure out. And I do think there's an, there's a big opportunity. We have a lot of free users there. I think how can we convert some of those into being willing to pay for, you know, some paid version of the product that still delivers on, on their needs?

[A]

If I recall when you guys brought in consumption based pricing to a degree, you said that you had two options. One was to not offer the latest and greatest models because they didn't fit into your pricing tier or to essentially give people the option to pay more for I think AI credits essentially to approach this. So would it be possible to kind of like go back to the SaaS approach and just have a much more limited selection of models that you could then sell at a price point that converts to a reasonable local equivalent in

[B]

India you definitely could. I think we are very mindful of like what's the experience we want our users to have? Like do, do we still feel like that will give them enough, a good enough experience where they can really depend on Gamma to do, you know, important work for them if they're, you know, high stakes presentation, does it deliver on that? And so I think that's more of what we're trying to balance out is making sure that we don't degrade the experience by either, you know, constraining the model selection to a point where the product itself doesn't, is not as usable as we'd like it to be. So obviously moving target. I think we can offer way, way more than we could ever. And you know, there'd be a great world where it would be an awesome place where, you know, maybe for most of your usage it's free and occasionally you can then actually lean into like some sort of usage based spike because you need that, you know, you need the most powerful thing for this one specific task. I think we'd be okay with that too. And so how can we actually have the flexibility? Is. Is ultimately where we want to go.

[A]

So in other words, we don't have enough compute, things are too expensive to run and you really need StarCloud to kind of, you know, get with the program here and get those GPUs in space.

[B]

Excited for that? Yes.

[A]

No, I just. There's going to come a day when I sit down with some folks like yourselves and they say we have so much compute, we don't even know what to do with it all. It's coming out of our ears and we are a long ways from that. But I do think once we do get there, it's going to be simply a fantastic time because you can build much more quickly. And I'll close on this with you, Grant. If you did have zero cost of compute, right. Let's say Philip wins. StarCloud takes over the whole sky is a data center. Unlimited inference. What would that do to Gamma? Would that make your company more defensible or less?

[B]

I mean, it's an impossible thought exercise because if you're pushing the boundaries of what you think is going to be possible, it'll come at a cost. So if we were just to stagnate in terms of, hey, just serve, what would the vast majority of PowerPoint decks look like today then? Yeah, then I think our market isn't very interesting. I don't think it would be interesting even if COMPUTE wasn't free. So you basically can ladder up to what's your next level of ambition and try to build the next thing. And so hopefully the costs do come down so we can continue to ladder up so that. But again, a few years from now, the way people are communicating just looks dramatically different and ultimately is in service of what they need to do.

[A]

No, I'm really hoping that you win because I feel like now that everyone's learned how to cheat with writing, writing has really fallen off a cliff. No one likes to do phone calls. We're all zoom burned out. So we need some new way to talk to one another before we bounce. I do want to take a look at a clip. This is everyone's dear friend, Sam Altman. He's on a podcast. He's talking about where we are today. And in his view, we are in the singularity, which I think we're going to need to define. And then I want to get everyone's opinion on if he's right or not. Take a listen.

[D]

And the thing that drives me the most is like, this is the most interesting, important thing I can imagine doing. And we are now like in the singularity. Like this is the moment. For the last 10 years ago, this was like a kind of far off dream at best. It seemed very improbable. And now we're like actually in the moment that we used to like talk about at the lunch table in a very not serious way. And so the thing that drives me is I've been waiting for this my whole life and I think it's going to be incredible, hugely positive, awesome for the world. I'm excited to get to work on that. I also think some of the alternative visions painted by other companies are like quite terrified and make sure that gets pushed against and is not what happens. But the main thing of what's different than 10 years ago is like, we're actually in it. Like this is the real. I think it is both true that it is all one crazy exponential and any one moment is not like the tipping point, and also that we are somehow in another one of those decisive periods where the curve can go one way or another like it was when we started 10 years ago.

[A]

All right, so Philip, starting with you, define the singularity for us and then weigh in if Sam's right.

[C]

Well, let me start on the second part. I'm 100% in agreement with Sam that we are in the singularity. To me, the singularity is. Well, typically it's just defined as this moment of recursive self improvement where there is this incredible acceleration because smart models create even smarter models which can create even smarter models which can create and so on. I like to think about it. Just imagine any kind of graph and where, if you were to zoom out, where does the graph go? And I think on any kind of like timescale, more than like a few years, when we zoom out historically, we have passed the point where the graph's gone. And so that graph could be like total GPU hour per year. It could be total energy going into AI per year. It could be total tokens produced per year. You know, any kind of graph you can imagine. Like, we are very much now past the point of no return. So I would very much agree that we're in singularity.

[A]

Grant, I'll go last because people want to hear from me the least. So why don't you wait here on yes and no and then definitions if you want.

[B]

My answer is yes as well. I mean it just depends on how you want to define it. Whether it's super intelligence, whether it's self improvement. I think it ends up becoming this sort of spectrum. And if you think we're in this sort of there is this period of exponential, then I don't think there's any denying that we're definitely in it. And then the question is what comes next? And I guess we'll have to wait and see.

[A]

The way that I think about singularity in this context is when do machines create net new intelligence faster than humans? That's kind of my working, definitely AI realm. And I think that we've seen now enough posts from different labs saying sparks of AGI and oh, we're seeing hints of recursive improvement here, things like auto research that we mentioned earlier and faster improvement of smaller models and such. The bubbles you want to see in the pot of singularity as you cook it. And so I definitely think we're there. And another example of this to the chart point is if you go and look at openrouter's rankings chart, it's on their dashboards on the website and you go back six months and look at the total number of tokens they were serving and you go back six more months and then you realize that they're adding more tokens per week now than they had entirely served like a year ago. And that is the pace that I think really drives this stuff forward. So, you know, yes, we need more compute. Yes, we haven't sorted out how to handle open models very well yet. But yeah, we're in it, I guess. Then the right question is, Philip, when does it get weird? Because right now my life is still pretty much the same with some cool tools and I can do more, but I'm still like mostly living my life the way that I was.

[C]

It's a great question. I think it gets weird when robotics takes off. When, when, when you, when you walk down the street and there's a humanoid carrying, carrying somebody's bag, I think that's when it gets weighed.

[A]

Grant, what do you think?

[B]

Yeah, I like that. I think that and when self driving cars become the dominant form of personal transportation, maybe my commute will get a lot better. So I'm hoping for that to come sooner rather than later.

[A]

It's never going to happen in San Francisco. It's never going to happen. I love you, but no.

[C]

Even if all cars were self driving, I Don't think I would find that that weird, to be honest. I think I'd be like, okay, now cars are self driving. Because when I get in a Waymo or a Tesla, it's just like that's the thing now.

[A]

So do you guys go back in time to the first time you used ChatGPT, right? Do you remember that, that little moment of magic when it did something that you found shocking as a capability and you were blown away? For me was I had it write funny little poems for my spouse and I was just making her laugh. And at the time that it felt like I had literally captured a genie in a magic bottle. That's how it felt now. Now she doesn't even like it when I use voice mode around her because she finds it annoying. And that's how fast that hype cycle went. Waymo is a great example, man. If you went back five, 10 years and said, by the way, if you live in Phoenix or San Francisco or LA or Miami and soon to be New York, you can get an actual self driving car, give them money and they'll drive you. People would have been minds blown. But now it's relatively pedestrian. Same thing with the Turing test. We blew by that. Like we didn't even notice it. And so I wonder if things actually won't get weird. I wonder if actually we're going to be really well adapted.

[C]

And another answer to that is it could get like absurdly weird, absurdly quickly. And by weird, I mean like a pdum type scenario. Like instead of like hugging face, shutting down a server, like the AIs actually just like take over everything all at once and it's a planned, coordinated thing and suddenly the world is like very weird very quickly.

[A]

You know, my PDIMM was so low up until this recent OpenAI agent breaking out of the sandbox, hacking, hugging your face and creating an actual problem. And the reason why it scares me is it took a long time for people to figure out what was going on and then to kind of back work their way back to figure out the issues and security flaws. That's way past a human's ability to stay on top of things. We're then dependent on AI systems and agents to solve AI systems and agents. And that makes me feel a little more p doomy.

[C]

My p doom is like 99.9% bill.

[A]

Why are you building compute then? You're. That's like sharpening your enemy's sword. What are you doing? Are you serious?

[C]

So it's, it's not so it's it's more like a philosophical point about the Fermi Paradox. Like, I, I'm still baffled that we don't see more intelligent life in the galaxy. And so I'm not necessarily saying it's going to be AI. It could be many different things. But it does seem that intelligent life in our galaxy at least is quite short lived for some reason. I'm not sure 100% why that is,

[A]

but yeah, the Great Filter. Or are we early? There's a lot of answers to the Hooray Paradox. Or just like, are we listening to the right things? Would life look like what we think we know? Grant, I'm really curious about your pdum. I know you're not exactly building software that's going to take over the world, but maybe you'll teach AI how to lie to us in this visual intelligence world.

[B]

Yeah, I feel like I'm maybe just too optimistic about everything. So I appreciate you guys bringing me back to Earth on some of this stuff.

[A]

All right, well, listen guys, thank you so much as always. I love talking to you. I love drifting through these questions and issues quickly. Before we go, where can people find your company online? And is there a job you're looking to hire for? You want our shout out to the audience in case that right person's listening? And Philip will start with you.

[C]

Yeah, we're, I'm very active on excellent info and starcloud.comcareers is where people can find jobs and we are hiring across the board. Mechanical engineering, electrical engineering, power electronics, thermal software and then also some. Now we're looking for, you know, people who've worked in data centers on the business development side of things.

[A]

So everybody and then some love that. Grant, tell people where they can find Gamma. And also, is there a job you want to shout out?

[B]

Yeah, we're on X and LinkedIn. I'm Grant Lee. You can look up Gamma as well. Careers. Gamma app. We're hiring across the board as well. So engineering go to market. We're building out a New York and London team. So if you're international, would love to connect as well.

[A]

Fantastic. All right, my name is Alex. This is this week in AI. We'll see you guys next week. In the meantime, keep prompting. Bye.