Inside the battles determining the AI winners and losers

AI is often discussed as a monolith. Either service providers, datacentre operators, frontier model developers, and chipmakers all make money and the buildout continues apace, or they don't and the buildout slows. For macro investors, maybe that framing works, and all of the major players do want to see growing demand for AI reasoning. But for stockpickers, this framing ignores the medley of partnerships, negotiations, and knife fights which define competitive strategy in the space. A customer in one area may be a supplier in another and a rival in a third. 

To really understand the company fundamentals, look within AI, rather than just at AI. Imagine how things might look from the vantage point of an AI company CEO. Our perspective:

If you are Dario Amodei at Anthropic, your path to success is clear—you are playing to win. You have arguably the best frontier models, with a rapidly-growing enterprise base paying a real price for Claude. You're fine with AI agents, which execute tasks autonomously, so long as yours are the best. Your goal is to ensure that your frontier models capture the biggest slice of AI economics. 

Fortunately, staying at the frontier is getting increasingly expensive, and as long as you can outspend every upstart, you can keep them at bay. 

You need capital for that, which investors have so far been eager to provide. Unlike your arch-rival OpenAI, you now generate cash rather than burning it, and you've had the good sense to cultivate or pantomime a conscience. As the Good Actor, when you declare cheap open-weight models a threat, the right people listen. It helps that most such models are Chinese and thus easy to smear. Capture the regulator, stop competitors from emerging, and maintain your technical and economic lead over OpenAI. Just watch out for open-weight models grown on American soil.

If you are Satya Nadella at Microsoft, you have great strengths. Your operations generate tons of cash, your debt is rated safer than US Treasuries, and your market cap is over $3 trillion. So you can handle a money-throwing contest. You have over 500 data centres running already, you own a big chunk of OpenAI, and your cloud revenues are growing rapidly from an already-huge base. 

But if your strengths are great, so too are your threats. AI labs have come for legal software, insurance software, financial planning software, even game development software. Without a cutting-edge AI model of your own, how long before Microsoft Office is on the chopping block? 

Still, you aren't defenceless. With your enormous user base and infrastructure footprint, you could make cheap open-weight models available to the masses, robbing the frontier labs of the profits they need to threaten you (your OpenAI stake be damned). That would also keep your data centres under heavy load, making them more profitable. Autonomous AI agents could boost your infrastructure business but complicate your software business. 

Think too of your employees, who are paid largely in shares, and your shareholders, who used to think you could grow without much capital investment. Capital intensity you can manage, partly through deals with neoclouds, the smaller data centre operators like Nebius. Nebius gets recurring revenue on attractive terms, while you get to keep assets off your balance sheet. If only you were further along with your own chip designs, you could lower your costs and avoid padding Nvidia's profits.

If you are Jensen Huang at Nvidia, that's exactly what you don't want. You reap astonishingly high margins from the chips you design, and to keep doing so, you need to keep being the best and you need to deliver adequate supply. 

Nobody likes paying your prices, but most have no choice—your chips are oppressively useful. For your part, you hate paying astonishingly high margins to memory makers like SK Hynix, but you have little choice for now, you can afford it better than most, and your designers are beavering away to make your chips less constrained by memory. You like open-weight models, which make it easier for users to afford AI reasoning and demand more of it, supporting demand for your chips. 

Autonomous agents would tilt the balance away from your graphics processing units (GPUs, good for reasoning) and towards central processing units (CPUs, good for managing and executing tasks), where your competitors AMD and Intel are closer behind. But you have great CPUs too, and only yours can be co-designed for synergies with your excellent GPUs. 

Helpfully, your business needs little capital, so you can throw cash around strategically. Deals with neoclouds and frontier labs keep your customers fragmented to reduce their negotiating leverage, while deals with suppliers limit capacity for your competitors. Plus you get unmatched visibility into the workings of cloud providers and frontier labs alike. This makes you the leading leather-jacketed warrior against bottlenecks. As long as growth continues, you can continue to thrive.

If you lead the memory makers SK Hynix, Samsung Electronics, or Micron Technology, you want to keep the boom times rolling. A few years ago, traditional memory (DRAM) was in its worst downcycle since the global financial crisis. AI has saved you, creating a new category in high bandwidth memory (HBM). HBM is faster at talking to GPUs, and it takes you three times as many wafers to produce. 

For you, that is a good thing: as HBM has grown, it has stolen capacity from DRAM, making traditional memory extremely scarce and profitable. Your customers hate this, but they need your supply, so they grit their teeth to sign long-term pre-payment and price agreements on lucrative terms. Keep soaking your customers, and they will innovate to weaken you while funding nascent Chinese competitors. Invest to maintain adequate supply, and you risk overdoing it, leaving your plants underused and bleeding cash if demand slows. 

With only three memory makers—down from a dozen in decades past—this balance may be easier to strike, but both customers and governments are pushing you to build faster and share the wealth. Build you must, but at a careful pace.

If you are Sundar Pichai at Google, you are playing both to win and to not lose. 

Like Microsoft, your operations generate lots of cash, your balance sheet is strong, your market cap, user base, and datacentre footprint are all huge, and your large cloud business is growing quickly. Like Microsoft, you have a golden goose—in this case search advertising—to protect. Like Microsoft, your people are paid largely in shares, and your shareholders are not used to worrying about capital investments or burning cash. 

But you alone are vertically integrated: your own chips, your own frontier model, your own datacentres, your own services. This gives you a cost advantage, which you can use to starve other frontier labs of profits, and it gives you a marketing advantage, as you can push any new product to billions of users. Your challenge is to keep your models at the leading edge, while keeping shareholders onside and ensuring that AI adoption does not upset your toll on internet traffic. Here, agents are a threat. What happens to online advertising when humans aren't doing the browsing?

If you are C.C. Wei at Taiwan Semiconductor Manufacturing Company (TSMC), much of the above is moot. You don't care who has the best services, the best infrastructure, the best model, or the best chip designs. When it comes to actually manufacturing cutting-edge chips, essentially all roads lead to you—whether those chips are designed by Nvidia, AMD, Broadcom, Marvell, or Google. Chip manufacturing involves nearly 1,000 discrete steps, so even 99.9% reliability in each step is insufficient. Near-perfection is necessary, and only you at TSMC have near-perfected the manufacturing process. 

Your task is to invest enough to stop Intel and upstarts from getting orders, while earning enough to justify the investment and keeping prices reasonable enough to support the rapid growth of the industry. As chip manufacturer for everybody, you have the world's best insight into long-term chip demand, and your company has an admirable and longstanding culture of striking this balance fairly.

For most of these companies, competitive strategy matters as much to their prospects as overall industry growth. 

The buildout could slow while some of the businesses thrive, and the buildout could continue while some of the businesses suffer. We think deeply about the competitive dynamics. As investors, we must then pair this with views on the companies' valuations. 

On our estimates, TSMC and Nvidia trade at about 23 times this fiscal year's earnings, the same as the wider market despite much better fundamentals. Alphabet trades at 18 times and was recently punished for raising capital investment—a healthy sign that investors have not lost their heads. Samsung's preferred shares (the main ones we own) trade at just 4 times earnings. SK Square, a holding company for SK Hynix, trades at just 3 times. Nebius made little money last year, but trades at 8 times our estimate of 2027 earnings.

While those valuations look calm, enthusiasm for AI is rife, and the related shares have been volatile. That makes us cautious, and the Orbis Funds are underweight the most AI-sensitive sectors in global stockmarkets. Mindful of giddy sentiment and potentially peaking memory profits, we have trimmed SK Square aggressively and Samsung moderately over the past two months. But while broad sentiment may be bubbly, we would not say the companies we own are in a bubble. Fundamentally, they look attractive, and picking stocks within AI is not the same as simply betting on AI.

Managed Fund

Orbis Global Equity Fund

Global Shares

View