Gabriel Costa

Founder Managing Director Venture Partner Board Member | Macro Finance geek/AI orchestrator

[Gabriel Costa @gpenacosta]

90’ in the match and Argentina hardly touched the ball 😅🇪🇸

[Gabriel Costa @gpenacosta]

accurate 😂

[Gabriel Costa @gpenacosta]

https://t.co/1GZzpXZZnA

[Happy Punch @HappyPunch]

The world’s first robot MMA event just went down in China 🤯

Kicked his head clean off 💀

[Gabriel Costa @gpenacosta]

Shenzhen just held the world’s first humanoid robot MMA event.

[Gabriel Costa @gpenacosta]

the key questions is:

is each incremental AI capex dollar becoming contracted high-margin cloud revenue, or is it becoming depreciating hardware used to subsidize low-ARPU inference?

the SpaceX deals are the cleanest expression of what hyperscalers want AI capex to become - AI infra as a contracted yield asset.

[Gabriel Costa @gpenacosta]

yes, but imo this dynamic holds only while the market believes the bottleneck is supply, not unit economics.

capex up = bottlenecks win because scarcity pricing/backlog.

capex down = hyperscalers win only if it reads as discipline.

the real bear case is not “AI demand dies” - i'm conviced ai demand is infinite.

bear case it’s that token demand explodes while value per token compresses, models get more efficient, routing pushes average tasks to cheaper inference, open/source/specialized models collapse the price of "good-enough" intelligence.

then the issue is not usage

it’s whether we overbuilt for the wrong token economy

fact is ai can be transformative and still produce bad capex cycles.

so the trade is not simple bull/bear AI.

imo next shift si more about a dispersion: own the layers where scarcity persists, be careful with the layers priced as if today’s scarcity is permanent

[Gabriel Costa @gpenacosta]

Average intelligence becomes cheap.

That may be one of the most important questions for the AI trade.

The market has mostly priced AI as a simple equation:

more AI usage → more tokens → more compute → more data centers → more semis → more revenue

But what if that chain is too linear?

What if token demand explodes, but the value per token collapses?

AI demand may explode. The problem is that the market may be underwriting the wrong token economics.

What if models get much more efficient?

What if routing pushes most tasks to cheaper models?

What if open-source and specialized models compress the price of “good enough” intelligence?

What if we overbuild infrastructure for today’s token economy, while the business model shifts underneath it?

That is the risk.

Not that AI is bubble.

Not that demand disappears.

But that the market may be extrapolating the wrong unit economics.

In software, the scarce thing was distribution.

In cloud, the scarce thing was infrastructure.

In AI, we still don’t know what the durable scarce thing is.

Compute?
Data?
Distribution?
Workflow ownership?
Trust?
Enterprise integration?
Regulation?
Human attention?

The answer matters because the current capex cycle is enormous.

If intelligence gets cheaper faster than usage grows, some parts of the stack may be overearning today.

If models become more efficient, the winners may not be the companies selling the most raw compute.

If the average task can be handled by cheaper inference, value migrates away from generic capacity and toward the layers that own demand, workflow, and differentiated outcomes.

So I don’t think this is a clean bullish or bearish AI take.
It is a dispersion trade.

AI can be transformative and still create bad investments.

Token volumes can explode and still produce pricing pressure.

Capex can be rational for one company and excessive for the industry.

The question is not whether AI usage grows.

It is whether the market is building and valuing the right parts of the stack for a world where average intelligence becomes cheap.

full deep dive:
https://t.co/3uPIvpxZKw

[Gabriel Costa @gpenacosta]

My 2 cents on SpaceX valuation, with the obvious bias disclosure that I’ve been invested since 2023.

Yes, if you look at SpaceX through traditional valuation metrics, the numbers look absurd.

Depending on what revenue number you use, you can easily end up with a very high price/sales multiple. On that basis alone, the stock looks “overpriced.”

But I think that misses the point.

in finance, there are basically two kinds of assets:

1. Assets that behave like bonds
2. Assets that behave like options

A bond-like asset is valued mostly by discounting predictable cash flows. Stable business, visible margins, low-to-moderate growth, DCF framework.

An option-like asset is different. You are not just valuing today’s cash flows. You are valuing a probabilistic distribution of future outcomes — including very low-probability, very high-payoff scenarios.

That is the right framework for SpaceX.

The key assumption is not current price/sales.

The key assumption is TAM — total addressable market.

If you believe SpaceX is “just” a launch company + Starlink, then the valuation is hard to justify.

But if you believe SpaceX is building the infrastructure layer for the space economy, then the question changes completely.

Falcon 9 already gave them a massive lead. Starship will be the next unlock: lowering cost per kg to orbit in an order from 10-30x cheaper and making entirely new markets economically possible.

Starship flips the economics.

Falcon 9 dedicated ~$2,700–$3,500/kg (or ~$7k/kg rideshare effective). Starship targets <$100/kg (Musk has eyed ~$10/kg long-term) via:

- Full reusability (booster + ship)
- 5-10× payload (100-150 t vs ~20 t)
- Rapid reuse + high cadence - Cheap methalox + stainless steel manufacturing
10-30× cheaper per kg.

That’s the flywheel: lower cost → way more flights → more revenue/data/iteration → even lower costs and higher cadence.

Makes bulk mass to orbit (servers, radiators, solar) viable for orbital data centers instead of prohibitive. Looking at current financials using F9 numbers miss this step-change. These are OoM improvements.

Space-based data centers. Orbital compute. Space solar power. Massive satellite networks. Lunar infrastructure. Defense. Communications. Energy.

That is the convexity.

The market is not simply pricing current revenue. It is pricing the probability that SpaceX becomes the dominant infrastructure platform for a market that could eventually be measured in tens of trillions.

So the real question is not:

“What is the price/sales multiple?”

The real question is:

“What probability do you assign to SpaceX becoming the monopolistic platform layer for the next phase of the space economy, and how big is that economy?”

If that probability is very low, the valuation is insane

If that probability is meaningful, then SpaceX is less like buying a normal company — and more like buying a call option on the industrialization of space. You are pricing convexity and optionality.

Let’s say there is a 20% probability that SpaceX will own 40% of a future space economy with a $20T TAM.

The implied expected value of that story is 0.2x0.4x20t = $1.6t which is not far from current ipo pricing.

And in a world where AI, energy demand, compute, defense, robotics, and space are all converging at once, that option may be worth a lot more than traditional multiples can explain.

[Gabriel Costa @gpenacosta]

@RaminNasibov full throttle
out of this world
stunt
carmen san diego
leisure suit larry
secret of monkey island

[RT @RT_com]

Brazil is going viral with ‘running raves’

Sao Paulo’s PACETRONIK events mix group runs with electronic music and club energy

The running rave concept is already backed by ASICS Brasil

Footage: PACETRONIK

[Gabriel Costa @gpenacosta]

once near-AGI systems can perform frontier AI research, intelligence itself becomes scalable infrastructure. Instead of relying on a small pool of elite human researchers, labs could deploy millions of AI research agents to improve the next model generation.

that is the “oh shit” inflection in the hockey stick curve: AI progress stops moving at human institutional speed and starts moving at machine speed. Years of algorithmic progress could be compressed into months, and human-level AI could become only a brief transition point on the way to vastly superintelligent systems.

the limiting factors are obviously compute, energy, chips, raw materiais, refining, manufacturing... but the incentives are too large for those bottlenecks to remain static. Capital, markets, governments, and hyperscalers will all converge on them. If energy and compute are solved, which i believe they will, the recursive loop becomes extremely hard to slow down

[Gabriel Costa @gpenacosta]

@threadreaderapp unroll

[Gabriel Costa @gpenacosta]

A content-driven launch strategy for a consumer product we worked on recently:

1. Create and distribute free content about the problem to be solved. Add value upfront to your target customer persona. You will start building an audience while generating &amp; qualifying leads.

[Gabriel Costa @gpenacosta]

7. Engage your audience as a community, create a sense of anticipation and scarcity, activate your leads list and only then officially launch your product.

[Gabriel Costa @gpenacosta]

@paulg this concept has been around for ages... you pay more if it's hand-made

[Gabriel Costa @gpenacosta]

@kode11 @paulg this concept has been around for ages... you pay more if it's hand-made

[Gabriel Costa @gpenacosta]

wild

[NewsForce @Newsforce]

POKÉMON GO PLAYERS TRAINED 30 BILLION IMAGE AI MAP

Niantic says photos and scans collected through Pokémon Go and its AR apps have produced a massive dataset of more than 30 billion real-world images.

The company is now using that data to power visual navigation for delivery robots, letting them identify exact locations on city streets without relying on GPS.

Source: NewsForce

[Heinrich @arscontexta]

https://t.co/JnIR1JG33B

[Gabriel Costa @gpenacosta]

love when claude does this - no buddy, you're doing this right now

[Gabriel Costa @gpenacosta]

love seeing this

history in the making, paulg knew it could be massive

[Niko Ludwig @Collateral_com]

Brian Armstrong's original Coinbase pitch (2012)

Listen closely and you can hear Paul Graham giving pointers on the storytelling