News Skalata is now Tall Order
About Team Portfolio Founders Investors AI Ops Writing Contact Pitch us
A young woman with a backpack reaches up in the dark to join glowing stars into a rising line
Perspectives

Five questions to ask yourself before investing in AI

Howard Marks says AI is the least predictable thing investors have faced in a generation. His new memo ends with five questions to ask before backing any of it, and we’ve answered them for seed.

·15 min read
Contents —
Contents
0%
Back to top

Howard Marks is one of the most respected investors alive. He co-founded Oaktree Capital and has written memos to clients since 1990. His January 2000 memo, bubble.com, warned about the dot-com crash weeks before it began.1

Warren Buffett has said his memos are the first thing he opens and reads.2

His latest is A Layman Thinks About AI.3

What the memo argues

  • AI is "the most consequential and least predictable development for investors in at least a generation." Who profits is close to unknowable.
  • Making money from it means being right about a lot at once. Make four predictions, each 60% likely to be right, and your odds of getting all four are about 13%.
  • Markets now lean hard on AI. A group of 42 AI-related companies drove about two-thirds of the S&P 500’s gains since January 2024.3
  • Intelligent investing is largely about price, and a less predictable bet can be the better one if it’s cheap enough.
  • It ends with five questions to ask yourself before putting money into AI.

The five questions

Marks says these questions work for any investment, AI or not. They’re arguably more important now, because so much about AI is uncertain. We’ve answered each one for seed, and compared it with the usual suspects (the labs, chips and data centres).

1. How many things have to go right for me to earn an acceptable return?

Buy shares in Nvidia or Microsoft and you need to be right about the technology, the demand for compute, the competition, the financing and the timing, and then whether the share price has already counted all of it.

A later-stage (Series B, C or D) round in an AI company comes with a different list that’s just as long. The company has to grow into a valuation that already assumes it will succeed, and then an IPO or a buyer has to turn up on time. Canva arguably missed that boat.4

For a seed fund, only a few things have to go right. It backs dozens of companies whose fates mostly don’t depend on each other, and one or two big outcomes can return the whole fund.

That’s because of where we get in. Say we invest at a $5 million valuation and the company later sells for $500 million. Our stake is then worth a hundred times what we paid, before dilution. If the fund is $50 million and still owns 10% at the sale, that one exit pays back the whole fund. A $1 billion fund would need a $10 billion sale to do the same.

Now try making a hundred times your money on Anthropic. It lists this month at a reported US$1.4 trillion,5 so it would have to become a US$140 trillion company. It could happen, but the big multiple has already gone to the people who backed it early. When Anthropic raised its first major round in 2021, it was valued at about US$620 million.6 It’s now worth more than 2,000 times that.

What has to go right at seed is mostly the founder

For any one company, building the product is rarely the hard part anymore, because a small team can now ship working AI software in weeks. What has to go right is getting customers to adopt it, and to keep paying once they have. The founders who manage it have usually spent years inside that industry first. They understand the problem first-hand, and they already have distribution, because they know the people who’d buy and can sign their first customers through their own networks. Distribution is now a moat.

Declan Vanderhor spent years as a geotechnical engineer filling in borehole logs on paper. Now TabLogs lets field teams capture that data on site, and its AI turns it into finished bore logs and reports with no retyping. When we first invested, Declan was the whole team and revenue was about $50,000 a year. It now brings in about A$3.5 million a year, with customers in 20 countries.

Assaf Stizki ran restaurants for more than 15 years before he co-founded Restoke, whose AI runs the back-of-house admin for restaurant groups in Australia and the US.

Founders like these know things about their industry that no outsider does, like why the old, awful process has survived for so long and who has to sign off on replacing it. That knowledge can be tested. Spend time with the founder and talk to the people they sell to, and you find out pretty quickly where the pain is. Try doing that with next year’s demand for compute...

2. If my central assumptions prove wrong, how much could I lose?

In AI infrastructure, the losses could be big and arrive all at once. Marks spends part of his memo on what he calls circular transactions, where chipmakers, hyperscalers, labs and data-centre builders lend to and buy from each other. Sometimes the money goes from one company to another and straight back again.3 It’s the corporate version of two mates shouting each other the same round all night. Each of those deals gets counted as revenue, which has helped push valuations up. That leaves further to fall, and if one company gets into trouble, the others can go down with it.

At seed, we’ll be wrong a lot, and on a single company we can lose the whole cheque. A portfolio of 50 early-stage companies, though, has about a 97% chance of returning its capital and a 73% chance of doubling it.7

We also invest in tranches where we can, putting more in as a company proves itself and the risk comes down. We led three investments in Doohly, a Melbourne advertising platform, over two years, building our stake to about 15% before Canva bought it for A$30 million.8

3. Which risks am I being compensated for, and which am I simply assuming?

A late-stage round prices an AI company as if most of its risks have been dealt with, so the investor is compensated little for the ones that remain. At seed the risks are plain to see, and the price reflects them.

The risk our LPs back us to take is judging founders before anyone else has. The founders worth backing at seed rarely look like the last winner, and there’s no dataset yet for what the next one looks like.

They’re also building in the right place. A lot of the value in AI is likely to end up with the businesses building on top of the models, not the labs underneath them. The labs are in an arms race. Whenever one pulls ahead, its rivals tend to catch up, and once several models are equally good at the same job, they have to compete on price. For a model of the same quality, the cost of running it has been falling about tenfold a year.9 Not to mention the copying. Chinese labs have been training their own models on the answers American ones give, a shortcut called distillation, and it keeps them only a few months behind.10

That’s all a problem for the labs, which have to keep spending more to stay in front. It’s great news for the businesses that use them. Every price cut makes their products cheaper to run without making them any less useful to customers, so more of the value trickles down to those businesses.

The internet went the same way. In the late 1990s, telecom companies borrowed heavily to lay fibre-optic cable around the world, expecting to charge for the traffic. They laid far more than anyone needed, prices collapsed, and Global Crossing and others went bankrupt.11 Amazon and Google were built on the cheap bandwidth they left behind.

AI gets priced against the work

The businesses on top also get to charge differently. Software used to be priced against other software, at a few hundred dollars a seat. AI products that do part of someone’s job get priced against the job itself.

Take Paperly, which builds AI operations for schools. Most schools juggle five to 15 separate systems for admin, and office staff spend much of the day moving information between them. Paperly replaces those systems, so a school weighs it against the staff hours it frees up rather than against another software licence. Daniel Dawson, one of its founders, met the problem as a systems engineer inside a school.

Priced against labour, plenty of industries that looked too small for venture are now big enough, especially for a small fund that doesn’t need a unicorn to generate significant returns for LPs.

The risk we’re simply assuming

The labs went from unknown to market leaders in a few years, and a lab can take a startup’s customers without ever targeting its market. Canva spent 13 years building a moat, and one new Claude feature started draining it. Canva is now cutting 22% of its staff.12

So we ask what would be left of a company if a lab gave away the same AI feature tomorrow. H3D would still have its patents and three of the world’s top five hearing-aid makers as customers. SupplyScope’s agents know the compliance rules of retailers like Woolworths, Big W, JB Hi-Fi and Bunnings, and work inside the process suppliers use to get onto those shelves.13 A general-purpose model has neither.

4. How much optimism is already incorporated in the price?

The labs are about as unpredictable as AI gets, yet Marks notes they’re priced at a premium.3 Even if OpenAI and Anthropic stay on top, their prices already assume they will. The two took around 14% of all venture funding worldwide in 2025,14 and retail investors will soon have the chance to pour their own money in.

Smaller AI companies aren’t priced anywhere near them. AI businesses valued under $100 million have changed hands for a median of 4.8 times revenue, according to a new report from valuation firm Sherwood Australia, against roughly 27 times for Anthropic and 43 times for OpenAI in their private rounds.15

Seed is just as unpredictable, and it’s mostly priced for it. A seed company has no share price, and its round is negotiated rather than traded. Hot rounds do get bid up, but most seed companies haven’t drawn a crowd yet, and when we lead, we’re usually agreeing the price with the founder before anyone else has noticed.

In 1979 Buffett wrote that you "pay a very high price in the stock market for a cheery consensus."3 A seed company hasn’t got a consensus yet, because hardly anyone has heard of it. That’s even more true in Australia, where median seed valuations run about 50% below the US for the same calibre of company, yet the country ranks third in the world for unicorns per dollar invested and first for decacorns.16

Where optimism does creep in at seed is pedigree. Investors pay up for LinkedIn profiles with "ex-unicorn" in the headline and walk past founders who’ve spent a decade inside an unglamorous industry. Declan wasn’t part of the startup scene and didn’t know a single venture investor when we met him. Ten years ago that would have made building a software company very hard. Now outsiders like him have a real shot at building the biggest companies in their industries. They’re also cheaper to back, because so few investors are looking for them.

5. Does it amplify the risk I’m already carrying, or does it diversify it?

If you own an index fund, you already own a lot of AI, because a handful of AI companies now make up a big slice of the market. If you have a super fund, it’s already made the AI exposure decision for you.

You can buy the dip in Nvidia (which Jensen has kindly provided).17 Just know what you’re adding to. Chipmakers, hyperscalers and soon the labs all rise and fall with how much the world keeps spending on AI. Even SpaceX, which merged with xAI and listed in June in the biggest IPO on record,18 is now partly a compute landlord, with Anthropic paying it about US$1.25 billion a month for one data centre.19 Buying more of any of them adds to the same bet.

Seed spreads that bet, because a seed company lives or dies on whether its own customers keep paying, whatever happens to AI spending at large. Explore Science builds AI systems that carry out original research and sells them to scientists and universities, whose research budgets don’t move with Nvidia’s share price.

Seed also moves on its own timetable, with no daily share price to fall with the market. And companies now stay private far longer than they used to, so a seed investor owns them through the years when most of their value gets made.

Where that leaves us

All five questions come back to which parts of AI you’re suited to judge. Each layer asks you to get something different right, and each can be bought at very different prices.

The AI stack

What each layer of AI is betting on

Where we invest
The rest of the stack

What has to go right

Can you check it?

The giants

At seed

Tap a layer for detail, or switch to By entry price to see what a seed cheque buys in each one. Examples are illustrative, not a ranking or a recommendation.

Marks, who called the dot-com crash before it began, says intelligent investing is largely about price, and that a less predictable bet can be the better one if it’s cheap enough. It’s the rule we invest by. Every layer has giants a seed cheque can’t reach, and companies still young enough that it can. We pick by entry price rather than by layer, which is why most of our portfolio sits in applications, but not all of it.

At seed, what has to go right is mostly a founder who knows a problem better than anyone. They also need to have picked a market that’s worth the effort. Both can be tested in person, which no model trained on the past can do for you, so a partner makes that judgement on every deal we do.

The price of a seed company hasn’t caught up with what it could become. And when one works, a single sale can return a small fund many times over.

Like Marks, we can’t forecast what an AI company will earn five years out. But adoption is still early. Anthropic’s own data shows that in about half of all jobs, fewer than a quarter of the tasks have been done with Claude.20

Platform shifts tend to turn a few tiny companies into giants, and the ones that do it for AI are being founded now, at a price a first cheque can still afford.

On whether the trillions going into AI infrastructure will pay off, Marks borrows a line from the economist Stuart Chase.3

For those who believe, no proof is necessary.
For those who don’t believe, no proof is possible.
Stuart Chase

Seed is the one part of AI where we don’t have to take it on faith. We’ve been backing founders at the first cheque for more than a decade, across more than 85 investments, and it’s the stage we’re built to judge.

We’re raising Fund III Our AI-native seed fund, open to wholesale investors.
Request the investor pack

Sources

Show all 20 sources
  1. Howard Marks, ‘bubble.com’, Oaktree Capital memo, 2 January 2000
  2. Columbia University Press, The Most Important Thing (Warren Buffett’s endorsement)
  3. Howard Marks, ‘A Layman Thinks About AI’, Oaktree Capital memo, 11 November 2026 (citing J.P. Morgan on the S&P 500, Buffett in Forbes, 1979, and Stuart Chase)
  4. AFR, ‘One model release from oblivion: Canva and Atlassian’s AI challenge’, 25 May 2026
  5. AFR, Chanticleer, ‘Howard Marks just nailed the only question that matters for AI investors’, 15 November 2026
  6. Dealroom, ‘Anthropic’s $124M Series A came almost entirely from tech founders, not VCs’ (citing Forge Global secondary data), 16 August 2026
  7. Simeon Simeonov, ‘Patterns of Successful Angel Investing’, 2011
  8. AFR, ‘Canva splashes $30m on Melbourne advertising start-up’, 25 March 2026
  9. Guido Appenzeller, ‘Welcome to LLMflation: LLM inference cost is going down fast’, Andreessen Horowitz, 12 November 2024
  10. The Economist, ‘China increasingly depends on copying American AI’, 25 October 2026
  11. CNN Money, ‘Global Crossing files for bankruptcy’, January 2002
  12. Canva, ‘An important update on our team’ (Melanie Perkins), 4 November 2026
  13. AFR, ‘Big box retailers turn to AI for back-office and compliance tasks’, 22 October 2026
  14. Crunchbase News, ‘The year in big funding trends, in charts’, 2025
  15. Capital Brief, Sweat Equity, ‘Figma of speech’ (Bronwen Clune, on Sherwood Australia’s AI valuation report), 18 November 2026
  16. Side Stage Ventures and Dealroom, Australian Venture & Startup Report 2026
  17. Bloomberg, ‘Nvidia’s stock drop made a tempting case to buy the dip’, 14 November 2026
  18. BNN Bloomberg, ‘After record IPO, Musk’s SpaceX faces next test in market debut’, 12 June 2026
  19. CNBC, ‘Anthropic, SpaceX announce compute deal that includes space development’, 6 May 2026
  20. Anthropic, ‘Anthropic Economic Index report: Learning curves’, 24 March 2026
Sam Henderson

Sam Henderson

Operating Partner · Tall Order

20 years of GTM and ops expertise across startups and ASX-listed tech companies.

← All writing