VC10X

Learn from world's top VCs, asset managers, family offices. Find published episodes at Capital Podcast by @ChoubeySahab

[Prashant Choubey @ChoubeySahab]

How to raise a VC fund - VC10X Playbook

https://t.co/YUU3TCxAA6

[Prashant Choubey @ChoubeySahab]

Access is now "table stakes" for top venture managers. So what do LPs actually test?

We asked 12 allocators, on the record. Their answers came down to 8 tests.

Free playbook + 40-question due diligence PDF: https://t.co/22R5vhVdrz

[Prashant Choubey @ChoubeySahab]

Everyone assumes AI is compressing the LP's job.

Tom Duffy, who allocates to venture funds at TIFF, sees it differently. Data is now commoditized. Everyone has the same information, the same screens, the same speed. So the thing that used to be an edge isn't one anymore.

What's left is the part machines can't do.

His phrase: they help you process more data and move quicker, but they don't replicate taste.

Which changes what he looks for in a manager. Not coverage. Not access to information. Focus, a genuine point of view on one part of the market, and the judgment to make a call on people that no tool will make for you.

What we cover:
• How TIFF underwrites a first time fund with no attributable track record
• The one word that separates the best managers from the rest
• Investing in AI through specialists rather than chasing the theme
• Why forcing exits to manufacture DPI can be a worse outcome for LPs
• Check sizes, explorer positions, and how an emerging manager becomes an anchor

New VC10X episode with Tom Duffy , Director of private markets at TIFF

[Prashant Choubey @ChoubeySahab]

The instinct that made Morgan Flager (@mflager) a good operator nearly made him a bad investor.

He started a company. He was an early employee at several others. They mostly worked.

That builds a specific reflex: give me a situation, however broken, and I can shape it into something that succeeds.

Carried into venture, it becomes a trap. You look at a disadvantaged company and immediately start listing the fixes. Change the strategy. Hire that person in six months. Solve this problem.

His correction, on VC10X:

"you can't apply the same ergs of energy to fifteen companies that you can to one"

And the sharper version. If you're smarter than the entrepreneurs across your entire portfolio, you either backed the wrong entrepreneurs, think too much of yourself, or both.

So you assess a company on what happens if it's left to its own devices.

But he immediately guarded against overcorrecting into pessimism, and this is the line that stayed with me. The most important attribute you can maintain as a venture investor:

"But what if it goes right?"

We also covered:

- Why Silverton Partners is at least doubling its hard tech allocation, to 30 to 40% of the fund
- What still makes a software company defensible, and why AI wrappers are worse than features
- Why he's writing more seed checks instead of holding deeper reserves
- The math that makes him walk away: a $300M seed entry needs a $5B outcome
- How LP conversations changed between Fund VII and Fund VIII
- Twenty years at Silverton Partners. Thirteen boards. Still asking the optimistic question.

Full conversation in the comments.

[Prashant Choubey @ChoubeySahab]

"I can't eat IRR."

Ned Brines (@nedbrines) said this about ten minutes into our conversation and I have not stopped thinking about it.

He runs the portfolios for a single family office, and when he underwrites a private fund he is not looking at the number on the front page of the deck. He is looking at what actually came back.

Because IRR gets manipulated, and the easiest lever is the subscription line. The manager uses a credit facility for an acquisition, calls capital six months later, and the IRR improves without the multiple changing at all. Same deal, same outcome, better headline number.

So his team disaggregates the LP cash flows, ties every capital call back to the actual deal date, and then checks what almost nobody checks: what was the last carrying mark before exit, versus what the market actually paid.

Topics we cover:

→ Why owning thirty managers is one bet wearing thirty different labels
→ How should families approach real estate investing
→ How IRR gets manipulated with subscription lines, and how to strip it back out
→ Why he caps venture funds at $150M and PE funds at $750M
→ Why he holds cash and waits for volatility instead of deploying into it

& lots more

Full episode link in replies 👇

#familyoffice #managerselection #realestate

[Prashant Choubey @ChoubeySahab]

Ken Goldman (@ken_goldman) ran Eric Schmidt's family office.

At one point, his team ran an analysis suggesting they could make more money investing outside of Alphabet than by holding it.

They never acted on it.

Alphabet is still, by a wide margin, the single best investment Schmidt has ever made.

Ken's takeaway was not about the analysis being wrong. It was simpler than that. People sometimes really don't understand how technology works.

He was CFO of Siebel Systems, then CFO of Yahoo, then spent five years as President of Hillspire when it already had more than 500 employees. Investments, legal, HR, IT, aviation, foundations, and an oceanographic research vessel, all in house.

We covered:

→ How you hire and retain when you have zero equity to offer
→ Why Hillspire kept Eric's personal bets separate from the institutional book
→ What makes a family office conference worth attending, and what makes you the product
→ Why he says we're in a bubble even though every insider he asks says AI is under-hyped

New episode of VC10X out now. (link in replies 👇)

[Prashant Choubey @ChoubeySahab]

A 2020 NVIDIA GPU just got booked through 2029.

That line from CoreWeave's Q2 call is doing a lot of work right now. Mike Intrator cited a recent A100 contract priced out to 2029. The CFO added that ASPs on older generations are sitting at or above where they were a year ago.

The GPU depreciation bears got quoted back at themselves within the hour.

I think that read is too fast. The same CFO said the renewal cohort is a very limited part of the fleet.

One contract on a small base proves old silicon still finds a buyer. Nobody serious argued otherwise. The question was never whether an A100 still runs in 2029.

It is what the box earns in year seven, and whether that clears the capital and the interest sitting behind it. Duration without pricing does not settle that.

The interest part is getting harder to look past.

Net interest expense hit $640M against $267M a year ago. Roughly $35B of debt sits on the balance sheet. Net loss widened to $626M from $290M.

The operating business is still compounding hard. Revenue came in at $2.58B, up 112% YoY and 24% sequentially, with adjusted EBITDA near $1.51B at a 59% margin.

Then the guidance.

FY26 revenue went to $12.4 to $13.2B from $12 to $13B. Capex went to $35 to $39B from $31 to $35B. Roughly $300M more revenue at the midpoint against $4B more spend.

Plenty of that capex funds 2027 and later, and a backlog near $104B plus the $25B+ signed early in Q3 gives it a reason to exist. Still, the ratio is the business model stated plainly.

The line I keep coming back to is inference. Managed inference ARR went from $1M to over $100M in a single quarter. The target is $250M by year end.

That is a different revenue shape than long dated training contracts, and a different customer set.

Shares rose 12 to 16% after hours. The market priced the growth, not the question.

What ASP would a 2029 dated A100 need to carry for that contract to clear the interest on the debt that bought it?

[Prashant Choubey @ChoubeySahab]

LPs cannot tell most funds apart.

I was reading the Q1 fundraising data and one number stopped me. 73.1% of all LP capital committed in the quarter went to five venture firms. The next ten split 15.4%. Everyone else fought over what was left.

That made me curious about what firms outside the top five are doing about it.

On paper the pitch is the same everywhere. Thesis, team, track record, a few logos. A $75M seed fund in Austin reads a lot like a $75M seed fund in Berlin.

Brand is what breaks the tie.

Not a logo or a color palette. Brand in venture means a clear answer to one question: what do you know that others don't, and who picks up when you call?

A podcast is the only format I know that proves both at once.

Every episode is a public record of your network.

You do not tell an LP you can reach a category-defining founder; they watch you do it.

It is also a live demo of how you think. Founders choose investors partly by listening to how someone asks questions. A podcast puts that on record, weekly, unedited.

Then there is the distribution math, where one recorded conversation becomes:

- a YouTube episode
- clips for LinkedIn and X
- a newsletter issue
- a searchable transcript
- a relationship with the guest

Five channels from one hour of work. Most firms run those channels separately and staff each one.

There is now a sixth channel that did not exist three years ago.

Founders and LPs increasingly start diligence inside ChatGPT and Claude. Who is good at seed-stage infrastructure? Which funds actually understand European go-to-market?

Ahrefs found YouTube mentions were the strongest single correlate of a brand appearing in AI answers, because the models read transcripts.

A podcast manufactures exactly that kind of source. Real conversation, real insight, your name attached, transcribed and indexed. Correlation is not proof, but the direction is clear enough to act on.

Meanwhile, most firms publish nothing a model can quote. A founder dinner is great for networking, but has no footprint on the internet.

Harry Stebbings closed a $400M fund for 20VC in four months, when the median US venture fund took over 15 months to raise. His LPs included MIT's investment office and RIT Capital. He was blunt about the reason: the fund is tied to a media business.

The podcast did not raise the fund. Track record does that.

But it compressed the timeline, and made the firm legible to people who had never met him.

Here is the part that should bother GPs. Brand compounds slowly and cannot be bought late. A firm that starts publishing during Fund I has a five-year head start by the time it markets Fund III.

Capital keeps concentrating, but attention has not consolidated yet.

Check our podcast production services for VCs at https://t.co/SiBZC7B0kb

[Prashant Choubey @ChoubeySahab]

There are many ways to build a successful business. Some are just more popular than others.

A founder mentioned Veeva to me last week while we were arguing about round sizes. I went and looked up the actual numbers.

Veeva raised $7 million in venture capital before its 2013 IPO. It spent only $3 million of it. The company listed at a $4.4 billion valuation.

Emergence Capital's $4 million Series A turned into a stake worth more than $1.2 billion. Roughly 300x, on one check.

For years that got filed away as a fluke.

It is turning into a category. Midjourney has taken no outside capital at all, and third-party estimates put its revenue near $500 million with a team of about a hundred people.

No board. No preference stack. No clock running.

What changed is not founder discipline, it is the cost of getting to first revenue. Stripe looked at the top 100 AI companies on its platform and found the median time to $1 million in annualised revenue was 11.5 months, four months faster than the strongest SaaS cohorts before it.

Engineering, support and go to market all got cheaper inside three years.

So the first round buys more ground than it used to. Sometimes it buys all the ground you need.

Founders now choose between three live options:

- One and done. Raise once, reach profitability, keep control and a clean cap table.
- Bootstrap. Slower start, zero dilution, full optionality on how and when you exit.
- The full ladder. Still the right answer for frontier models, chips, robotics, anything with real physical cost.

Here is the part allocators should sit with.

Capital is concentrating at the exact moment capital is becoming less necessary. In the first half of 2026, OpenAI and Anthropic together absorbed close to 43% of all global startup funding, while seed stayed small enough to look like a rounding error.

If a great application-layer company only needs $5 million, a $2 billion fund cannot own enough of it to move the fund.

That math pushes large funds toward the capital-hungry end of the market. Which is also the end with the highest entry prices and the least certain outcomes.

Fund size stopped being a positioning choice and became the strategy.

The uncomfortable read: the next Veeva is being built right now. Most of the industry is structurally unable to own enough of it to care.

[Prashant Choubey @ChoubeySahab]

Free weights do not mean free compute.

A founder told me his inference bill barely moved after switching to an open model. That made me curious, so I spent a few hours working out where the cost actually sits.

Most of the price gap has nothing to do with being open.

It comes from architecture. DeepSeek-V3 carries 671 billion parameters, but only about 37 billion activate on any given token. That is mixture-of-experts: you hold the full model in memory and pay compute only on the slice that fires.

Closed labs use the same design. They just do not have to pass the saving on.

The second reason is margin.

When weights are public, nobody earns rent on the model layer. Serving becomes a commodity trade fought on utilisation and hardware. Frontier API pricing carries research recovery and margin on top. Open serving mostly does not.

The third reason is less flattering. The cost did not disappear, it moved onto your balance sheet.

You now own idle GPUs, ops headcount, failover and monitoring. Below roughly 60% sustained utilisation the arithmetic stops working. Most real workloads are spiky, not steady.

So the cheapness is real, but conditional.

I assumed the capability gap had closed by now. Stanford's 2026 AI Index puts the top closed model 3.3% ahead of the top open model as of March 2026, wider than the 0.5% gap in August 2024. Epoch AI measures the lag at about four months through the first half of this year.

Four months is noise for a document classifier. It is not noise where a wrong answer creates legal exposure.

So nobody switches wholesale, they route, and routing is where the compute story turns.

Cheaper inference does not reduce usage. It changes what is worth running at all.

- Work too expensive to automate last year becomes the default this year
- One prompt becomes an agent loop of a dozen calls
- Every call re-reads context and re-emits tokens

Per-token prices have fallen by orders of magnitude since 2023. Enterprise AI spending rose anyway, and inference has overtaken training as the bigger line item.

Jevons watched the same thing with coal in 1865, when better engines made coal worth burning in places it never was.

For allocators, the useful question is where the margin lands. Open weights compress it at the model layer and push it outward.

Toward inference providers competing on utilisation. Toward memory, power and interconnect. Toward the application layer that owns the workflow.

Open models will not need less compute. The question is who gets paid when they need more.

*not investment advice. DYOR.

[Prashant Choubey @ChoubeySahab]

Most heirs get the money before the reps.

It is the real succession problem. And it never shows up in the estate plan.

UBS surveyed 307 single-family offices this year. Average family net worth of $2.7 billion. The governance findings are worth sitting with.

- 60% have an investment committee
- 35% have a succession plan for the family office itself
- 27% have an organized process to prepare the next generation for future roles

So the portfolio is institutionalized. The people are improvised.

Most families read this as an education gap. They respond with a curriculum. A weekend workshop, a family constitution, a module on trust structures.

That misreads the problem.

Allocation judgment is not knowledge, it is a skill built by making calls, being wrong, and living with the consequence while someone more experienced watches.

You cannot teach that in a seminar. You can only hand someone a mandate small enough to survive and real enough to matter.

The data says most families do neither. UBS found 45% involve the next generation fully or partially in decisions. Another 21% have heirs old enough to participate and no involvement at all.

It starts earlier than people think. Cerulli found that 34% of high net worth heirs learned the details of their inheritance only after the benefactor had died.

Their first decision as a principal is also their first decision with full information.

The second mistake is what families optimize for. In last year's UBS report, 64% named tax-efficient transfer as their biggest succession challenge. Preparing the next generation came in at 43%.

Tax leakage is measurable. So it gets the attention. Judgment decay is not, so it gets deferred until it is somebody else's problem.

Then the bill arrives. Capgemini surveyed inheritors and found 81% intend to replace their parents' wealth management firm within a year or two of inheriting.

Calling that disloyalty misses the point, because inheriting a portfolio you never helped build and relationships you never chose is exactly the setup for a clean sweep.

Wealth transfers on a date. Judgment transfers over a decade.

Most families only plan for the first one.

[Prashant Choubey @ChoubeySahab]

Most family wealth moves sideways before it moves down.

Cerulli puts the US transfer at $124 trillion through 2048. UBS estimates roughly $83 trillion globally over the next two decades. The headline reads like a generational handover.

The sequencing says otherwise.

Of the money flowing to people rather than charity, about $54 trillion passes horizontally to a surviving spouse first. More than 95% of that goes to women. Roughly $40 trillion of it lands with widows in the Boomer cohort and older.

So the first transfer event in most families is not a 35-year-old with a private credit thesis. It is often a 78-year-old who was never in the room when the portfolio was built.

Then there is the question of what actually transfers. UBS surveyed 307 single-family offices in early 2026, average family net worth $2.7 billion. Of those families, 77% still run an active operating business.

Add alternatives at 42% of the average portfolio and the picture is clear. This is not a brokerage statement, it is an operating company, a property book and a stack of unfunded commitments.

Estate liabilities do not wait for any of that to become liquid.

The governance gap sits on top. Only 35% of those family offices have a defined succession plan for the office itself. Just 27% run a structured process to prepare heirs for their roles.

These are not amateur shops. 68% have formal performance measurement. 60% run an investment committee. The machinery is built. The instructions for handing it over are not.

Three things follow for anyone managing this capital:

- Manager relationships are personal, not institutional. Cerulli has found that more than 70% of heirs change or drop the incumbent advisor. Twenty years of trust with the founder is worth very little to the successor.

- Secondaries supply is structural, not cyclical. Liquidity needs around estate events recur regardless of where marks sit.

- Near-term flow is smaller than the headline. Gen X inherits roughly $14 trillion over the next decade, Millennials about $8 trillion. Peak transfer is projected for the mid-2030s.

One caveat worth stating plainly. The range on these estimates is wide, and some models put transferable wealth far below $124 trillion once retirement spending, debt and taxes are netted out. Direction is not in dispute. Magnitude is.

[Prashant Choubey @ChoubeySahab]

Venture Capital Goes All In on AI.

A year ago, just under half of global venture dollars went to AI companies. In Q2 2026 that crossed 70%, per Crunchbase. PitchBook's US-only count puts it at 86%.

Roughly a 20-point swing in four quarters.

But the sector mix is the smaller story. The bigger one is concentration.

OpenAI and Anthropic absorbed $217 billion in the first half of 2026. That is 43% of every venture dollar deployed globally. Anthropic's $65 billion round alone took close to a third of the second quarter.

Meanwhile deal count fell to its lowest level in a decade.

Record dollars, fewest deals in ten years. Same quarter.

The funnel shows it plainly:

- Global H1 2026 funding hit $510B, more than all of 2025 ($440B)
- Megadeals of $100M and above took 87.5% of US capital deployed
- In Q1, 31 US deals of $500M or more captured about 80% of US funding
- The 1,020 US deals between $1M and $10M split roughly 1%

Capital also moved out of software and into things you can touch. Robotics pulled $69.4 billion in H1, while defence, semiconductors and power infrastructure took most of the large non-AI rounds.

Fintech managed $12 billion in Q1. Digital health $7.4 billion. Numbers that would have led the table in 2019 and now read as rounding errors.

Then the part allocators should sit with.

Fundraising concentrated exactly the way deployment did. Andreessen Horowitz, Thrive and Founders Fund took 48.1% of all US venture capital raised in H1, and first-time fund formation is tracking its weakest year since 2016.

So the manager selection question changed shape. It is no longer mostly about who picks well. It is about who gets allocation.

One real improvement. Liquidity returned, with SpaceX listing at $1.77 trillion and 24 companies acquired above $1 billion in Q2 for $113 billion, the strongest exit quarter since 2021.

Worth flagging that Crunchbase, PitchBook and CB Insights run different methodologies and report different totals, though they agree on the shape.

What would change this read: AI's share of dollars drifting back toward 50% while deal count recovers. Or a frontier lab down round that resets late stage marks.

Until then, most venture exposure is a concentrated position in a handful of companies, whether or not that was the mandate.

[Prashant Choubey @ChoubeySahab]

Most family offices want the Sequoia logo.

Lara Nuchowicz (@LaraNucho) wants space on the cap table at Series A.

Those are not the same thing, and the gap between them is the entire reason her firm backs emerging managers instead of brand names.

She runs private markets at AR Capital. More than 80 funds in the portfolio. Almost none of them are names you'd recognize.

Her reasoning:

→ A brand name will not give a family office follow on access to its breakout companies
→ Big funds drift from their thesis because they become platforms
→ Small funds return bigger multiples, so she caps at $100M and prefers under $50M
→ She tracks talent before thesis, because nobody can predict the next thing
→ Most managers she has backed lately are under 30.

The line that stuck with me: "It's much tougher for family office or for us to go to Sequoia and ask them space on the cap table."

New @VC10X_ episode with Lara Nuchowicz, Principal & Next-Gen Allocator at AR Capital.

Link in replies 👇

[Prashant Choubey @ChoubeySahab]

Alphabet beat expectations, and the stock still fell.

That is the frame for this week. Microsoft and Meta report on Wednesday, Apple and Amazon on Thursday.

The market's reaction function has flipped. For three years a bigger capex number read as confidence. It now reads as risk.

Look at what Alphabet actually delivered on 22 July.

- Revenue of $119.8B, up 24%, ahead of consensus
- Google Cloud up 82% to $24.8B
- Cloud backlog of $514B, higher by roughly $50B in a single quarter

Then the other side of the ledger. Capex doubled to $44.9B and the 2026 guide moved to $195-205B. Free cash flow came in at negative $5.9B, the first negative quarter since the 2004 IPO.

The shares fell about 5%.

The funding mix is the part worth sitting with. Alphabet raised $49.6B in equity and convertibles in June, plus $20.3B in senior notes. Long term debt has more than doubled since December, to $98.2B.

Buybacks were zero, against $13.2B a year earlier. An equity story is quietly turning into a credit story.

The second issue is timing. The five largest spenders are tracking roughly $760B of capex in 2026 while recognising about $211B of depreciation against it. The rest sits in construction in progress, waiting to be placed in service.

Depreciation is capex on a delay. It flatters reported earnings now and pressures them later, whatever AI revenue does by then.

Third, a bigger capex number is not the same as more compute.

SemiAnalysis estimates memory alone will absorb close to 30% of hyperscaler capex this year, against roughly 8% in 2023 and 2024. Part of every raised guide is simply paying more for the same rack.

IBM gave the receipt on 14 July. Clients pulled forward server, storage and memory purchases in the final weeks of June to lock supply before price increases. Software deals slipped and the stock fell 25%, its worst day on record.

So three things worth watching, in order:

- Azure growth against the 39% to 40% constant currency guide, and the first real read on FY27 capex
- Meta's capex range, last set at $125-145B, and whether operating margin holds near 41%
- Amazon's free cash flow, which has been running close to zero on a trailing twelve-month basis

The bull case sits in one line. If the contracted backlog keeps growing faster than capex, the spend is chasing demand already signed. Alphabet's backlog did exactly that last week.

The bear case is the funding mix, not the demand. Cash goes out now and the earnings charge lands later, on borrowed money.

How much they spend is no longer the question. Who funds it is. So is when it lands on the income statement.

#AICapex #Hyperscalers #Markets

[Prashant Choubey @ChoubeySahab]

Google beat everything. The stock fell 7%.

Alphabet's Q2 revenue came in at $119.8B, up 24% and ahead of the roughly $117B consensus, with operating margin expanding 200 basis points to 34%.

Then the market got to the capex line. It sold.

Here is what each part of the business actually said.

Search and other did $63.3B, up 17%. The bear case for three years has been that AI answers would eat the query funnel and compress monetization. So far the numbers say the opposite.

Cloud is where the acceleration sits. Revenue of $24.8B, up 82%, against consensus near $22.4B, and that follows 63% growth in Q1 and 48% in Q4.

The margin is the more interesting part. Cloud operating income tripled to $8.8B. Segment margin moved from roughly 21% to about 36%, which is what scale finally showing up looks like.

Backlog reached $514B, up more than $50B in a single quarter. Management says roughly half converts to revenue inside 24 months.

YouTube was the soft spot. Ads grew 13% to $11.1B. It beat expectations, but it is now the slowest growing line in the house.

Other Bets stayed a cost line. Revenue of $382M against a $1.8B operating loss, wider than the $1.24B a year ago.

Now the number that actually moved the stock.

- Q2 capex: $44.9B, up 107% year on year
- FY26 capex guide: $195B to $205B, raised from $180B to $190B
- First half spend: $78.6B, implying $58B to $63B per quarter in H2
- Free cash flow: negative $5.9B

That last line matters more than the guide. Alphabet has been one of the most reliable cash machines in corporate history. This quarter it consumed cash.

It did that while holding $242.5B in cash and securities. In June it still issued $49.6B of equity and preferred stock, on top of $20.3B of notes raised during the quarter.

You do not raise capital like that with a balance sheet like that to bridge a single hard year. That looks like a multi year plan being disclosed through the financing statement rather than the call. Capex guidance has now gone up three quarters running, and 2027 is still described only as significantly higher.

One caveat on the headline: GAAP EPS of $9.11 is not earnings power. It carries a $98B net gain, mostly unrealized marks on holdings like SpaceX and Anthropic, against adjusted EPS of $2.85.

The question for allocators is not whether this business is compounding, because it clearly is.

It is whether you are still underwriting a capital light compounder or an infrastructure operator with an unusually good customer list. Those two things carry different multiples. And the depreciation on this year's spend has barely started hitting the P&L.

At roughly 21 times forward earnings with 24% revenue growth, the market has priced part of that shift. Not all of it.

#Alphabet #AIInfrastructure #PublicMarkets

*only for informational purposes, not investment advice. Do your own diligence.

[Prashant Choubey @ChoubeySahab]

The highest-demand role of the next decade might be the non-technical generalist.

As AI absorbs technical execution, the scarce input becomes judgment. Someone still has to decide which output to trust and which to throw out.

That call is not a technical skill, and it is getting harder to hire for.

It comes from taste. From context. From the pattern recognition you only build over years.

The data is starting to confirm the shift.

PwC's 2026 AI Jobs Barometer found entry-level roles most exposed to AI are now 7x more likely to demand senior human skills like leadership and creativity. Openings for them grew 35% since 2019, while other entry-level roles shrank 10%.

The pay is following the same line.

- Roles that pair AI with human judgment saw salaries climb 42% faster

- The wage premium for AI skills hit 62%, up from 57% a year earlier

- 39% of core work skills are expected to change by 2030, per the WEF

Here is what sits under the numbers. AI is cheap at generating options and expensive at knowing which one fits this customer, this market, this moment.

That gap is where generalists with a strong taste win.

They move across product, finance, legal, and operations, and translate between all of them. They rarely go deepest in any single silo.

Instead they orchestrate the specialists and the tools around a problem.

For allocators, this changes how you read a business. The edge is drifting from headcount and process toward the quality of decisions a small team can make.

So ask a sharper question of management.

Not how many engineers sit on the payroll. But who holds judgment, and how good it actually is.

Taste is hard to fake. Hard to copy. Slow to build.

In a world where everyone holds the same AI, that may be the last durable edge.

[Prashant Choubey @ChoubeySahab]

Every AI dollar flows through five layers.

Where that money pools, and where it leaks, has become the central question for anyone allocating to this theme. The numbers are worth sitting with.

Layer one is power, and it is now the binding constraint.

Goldman Sachs expects US data center power demand to more than double, from 31 GW in 2025 to 66 GW by 2027. Microsoft has an $80 billion Azure backlog it cannot fulfil, with GPUs sitting idle because the electricity is not there. Demand is not the problem.

Layer two is silicon.

Nvidia's data center segment did $75.2 billion in its latest quarter, up 92% year over year. Most of the stack's profit has pooled in this layer so far.

Layer three is the infrastructure buildout, and this is where the scale gets hard to grasp:

The big five hyperscalers have guided to roughly $650 to $750 billion of 2026 capex, up over 60% from a record 2025

About 75% of that is going into AI infrastructure

Capex intensity now runs from roughly 25% of revenue at Amazon to over 85% at Oracle

They raised over $100 billion of debt in 2025 to help fund it

Those are utility-style capital ratios, increasingly financed through debt markets. That changes the risk profile of businesses long valued as capital-light.

Layer four is the models, and this is where 2026 surprised everyone.

Anthropic went from a $9 billion run rate at the end of 2025 to $47 billion by May, overtaking OpenAI at roughly $25 billion.

Combined, the two leaders now run above $70 billion. Real money, but still a fraction of the near $700 billion being poured in beneath them.

Layer five is applications, still the thinnest layer by revenue.

That is the tension worth watching. The layers with the least revenue today are the ones that must eventually justify everything built below them.

For allocators, the practical takeaway is that "AI exposure" is not one trade.

Power, silicon, infrastructure, models and applications each carry different margins, different capital needs and different failure modes. Sizing them as a single theme misses the point.

*only for informational purposes, not investment advice. Do your own diligence

[VC10X @VC10X_]

Who dis?

[Prashant Choubey @ChoubeySahab]

Investing in next-gen marketplaces

@ColinGardiner shares the investment thesis at Yonder Ventures on @VC10X_🎙️

Catch the full episode on 'VC10X' YouTube, Spotify, Apple Podcasts (link in replies 👇)

#marketplace #startup