This time last year we had the DeepSeek Moment, which turned out to be a misunderstanding by generalist investors and journalists over a few details in a research paper. NVIDIA dropped -20% in a matter of days, and voices from every corner of the market declared the end of model scaling and major investments in AI infrastructure. This has proven to be almost comically wrong, with hyperscaler capex in 2025 realising about +50% higher than the street expected and the latest hyperscaler capex guides for 2026 nearly 3x higher than the expectations of 18 months ago.

There have been many other bearish narratives over the last three years. There is the question of ROI - is there enough to justify these investments? Maybe the utility of AI will be permanently handicapped by inherent flaws like hallucinations, context length, or lack of long term memory. After all, 95% of pilots fail (per flawed MIT study last August). The latest flavour of bearishness is being called the SaaSpocalypse, which has driven a sell off across the tech sector. 

It's been 15 years since Marc Andreesen's Why Software is Eating the World essay which aged pretty well up to now. Software is a broad church, encompassing all aspects of digitalisation from entreprise software to consumer apps to payment networks and financial services. It includes cybersecurity, industrial design software, and systems of record that underpin corporate functions across finance, operations, HR and IT. Software companies surfing the digitalisation wave over the last decade have been some of the best compounding revenue growth stories in the stock market, all the while delivering 70-80% gross margins. Companies like Salesforce and Workday benefitted from low churn, due to high friction associated with data movement. This Golden Age for SaaS was underpinned by the ZIRP era, and investors rewarded top-line revenue expansion over profitability, accepting high cash burn and massive stock-based compensation (SBC) packages. The 'Rule of 40' was frequently achieved entirely through growth, often with negative FCF margins.

A substantial chunk of the software sector outperformance has unwound over the last few weeks, with software stocks broadly down about -30% from the highs.

The Golden Age of SaaS: The global SaaS market grew from $31BN in 2015 to $297BN in 2025e - a +25% CAGR over a decade

Software has unwound all of its outperformance versus the S&P 500 since YE17

We are definitely Claude Code-pilled at Green Ash. We wrote about the Age of Agents in March 2025, and our early experiments vibe coding some deterministic workflow automations last summer. In the two weeks since writing about the Great Unhobbling, the whole investment team has Claude Code installed and uses it daily. 

But for all its capabilities, we won't be cancelling our $20k per annum Bloomberg subscriptions any time soon, even less so our $300 per annum Microsoft Office subscriptions. We will still have need for Bloomberg financial data and market news and some terminal features relating to electronic trading and IB chat. If we could unbundle some of these things, we would, but Bloomberg will never offer this, and that is their moat. Anything output by Claude Code will likely be in the form of a Word document, PowerPoint presentation or Excel spreadsheet  - we will still need to review these files, make edits, etc. That is Microsoft's moat. 

We can only speak for our experiences as a boutique asset manager, but we imagine that this is true across many types of knowledge work and sizes of firm. Anthropic themselves are said to use Salesforce/Slack, Workday, Github, Atlassian, and MongoDB internally. It just doesn't really make sense for everyone to independently vibe code 20 years of software development to reduce an IT budget which is relatively tiny versus something like headcount. The code could have errors that no-one can spot, and there would be no ongoing support for the codebase to ensure reliability. 

When we set out to write this piece, our conclusion was straightforward: AI is reshaping software, some jobs will be lost, but revenues won't collapse - just another overreaction like the DeepSeek Moment. But as we read and thought more, we came to realise the outlook is far more nuanced, with potentially serious consequences for software company valuations.

Value Capture and Services-as-Software (SaS)

For 15 years, the 'System of Record' (SoR) was the ultimate prize, protected by the theory of data gravity - the concept that as data accumulates, it builds a 'mass' that attracts applications, services, and additional data to it - much like a planet attracts objects - making it increasingly difficult, expensive, and latency-prone to move that data elsewhere. While we think SoRs will remain critically important, the risk for incumbents is that they become 'dumb pipes', as fewer humans interact with the software built on top, and agents simply access the data directly via API. There is far less vendor lock-in in this scenario - agents are agnostic to file types and data formats; they can parse them interchangeably and translate them into any output as easily as Google Translate can convert English to Mandarin. Humans using software were the value layer on top of the data: both may be replaced by agentic layers which use neither humans nor software. 

This is the Services-as-Software (SaS) scenario: Where in the traditional SaaS model, a vendor sells a tool to a human to make them more efficient, in the SaS model, the vendor sells the outcome of the work itself, with the software acting as the worker. The economic capability and value of SaS allows vendors to attack a much larger prize: the labour budget. Historically, software spend was roughly 5-10% of a department's total cost, with 90% going to human salaries. By replacing the human labour, SaS vendors can capture a portion of that 90%, significantly expanding their TAM. 

OpenAI is building five layers of value on top of systems of record

While the overall profit pool of software is forecast to grow, a significant proportion of traditional SaaS revenues are expected to be cannibalised by agents

This profit pool forecast could prove conservative. Software-as-a-Service (SaaS) may transition to Services-as-Software (SaS), shifting from the IT budget to the labour budget

70% of SaaS revenue ($210-245 billion) is per-seat/per-user. US knowledge workers number ~65-70 million (BLS), each generating $4,000-$6,000/year in SaaS spend (UBS). If AI reduces knowledge worker headcount by -10-20% over 5-7 years, $21-49 billion in annual per-seat SaaS revenue is at risk.

Tech Debt and Build or Buy

Gartner estimates global tech debt exceeds $5 trillion (including opportunity costs). Stripe's 2018 developer survey found developers spend ~33% of their time on tech debt, costing $85 billion annually in lost productivity. At legacy-heavy enterprises, 70-80% of IT budgets go to "keeping the lights on" vs. innovation (Accenture/BCG).

Our favourite example of the heavily tech indebted is the legacy banking sector. An estimated 220 billion lines of COBOL remain in production globally, processing 95% of ATM transactions and 80% of in-person transactions in the US. But the average COBOL programmer is 55-60 years old, creating an acute talent cliff. AI is making modernisation feasible for the first time: last summer, the WSJ ran a piece about Morgan Stanley building their own AI tool internally to review 9 million lines of legacy code in languages like COBOL and convert it into plain English specs that developers can then use to re-write it in modern programming languages (saving 280,000 developer hours). Given the progress in coding model capability since then, it isn't too hard to imagine whole codebases being fully fungible between programming languages at the press of a button. If banks can modernise their core systems with AI, they no longer need SaaS band-aids layered on top. 

If companies can use AI to refactor and transform their internal systems in this way, why wouldn't they just build their own software applications? For a decade, companies bought point solutions for every micro-problem, resulting in the average mid-sized corporation accumulating hundreds of SaaS subscriptions. It's quite clear to us from our own internal vibe coding efforts that a long tail of specialist services companies can be replaced, though we are a bit more sceptical about the very large incumbent platforms and business critical applications. Code generated by non-engineers often lacks structural integrity, documentation, or security best practices, and, at a certain scale, it is impractical to vet and maintain very large AI-generated codebases. 

Terminal Value

Reducing terminal growth from +3% to +1% while increasing WACC by +200bps (from 10% to 12%) implies a -32% decline in enterprise value - even with no change in near-term fundamentals

One illustrative example of this effect is PayPal. Riding high from the COVID-era, when cash use collapsed, the narrative focus changed to take-rate compression in payment networks and competition from digital wallets from Apple and Google. The company maintained top line growth over this period (albeit decelerating) and reduced their share count by -22%, yet the P/E multiple has compressed by -90% (along with the stock price). The bear case for software value capture is analogous to PayPal's plight - if the value is to accrue at new agentic layers, valuations will permanently adjust accordingly, and many software companies have significantly lower FCF margins than PayPal's 17%, especially after accounting for share-based compensation. 

A re-pricing of terminal value can be very painful for valuation multiples for a long period of time

Adobe and Salesforce were generally viewed as AI winners until 2024, when the narrative inverted. Multiples have contracted ever since, despite solid underlying fundamentals

FCF margins for software stocks are much lower than they appear when accounting for share-based compensation

Private Markets

Terminal value headwinds pose a serious risk to private capital funds which have steadily increased their software exposure in recent years. 

Private equity funds are already sitting on $3.2 trillion in unrealised Net Asset Value, spread across 29,000 unsold companies. Even without AI, the industry is suffering from investments made in the ZIRP era with negative IRRs in several vintages over that period. Software still accounted for ~40% of private market deal volume as recently as 2024, meaning the exposure is being compounded, not reduced. 

In private credit, 25-35% of loan books face risk of AI disruption (the upper end of the range includes shadow exposure, via healthcare or business services portfolio companies which are vertical SaaS platforms in disguise, e.g. electronic health records or billing software). Private credit software exposure amounts to $400-450BN - nearly a third of the size of the entire US public loan market. A lot of this debt came from private equity companies juicing returns by levering up portfolio companies, in order to pay distributions to investors. CLOs typically capped software exposure at 10-15%, so borrowers turned to private credit funds for liquidity (private credit lenders financed 93% of software LBOs in 2023). For some historical context on what can happen in times of rapidly worsening industry fundamentals, when the fracking industry blew up in 2015-16, sub-investment grade credit defaults reached nearly 30% for the E&P sector. 

The large firms are aware of this - at a gathering in Toronto in late 2025, John Zito of Apollo Global Management commented that "the real risk is software is dead". Apollo CEO Marc Rowan echoed this sentiment, stating that technological change will cause "massive dislocation in the credit market," adding on entreprise software, "I don't know whether that's going to benefit or be destroyed by this. As a lender, I'm not sure I want to be there to find out".

Business development loan companies (BDCs) have an average of 25% of their loan books invested in software companies

We asked Claude Code to model the fallout from -20% and -30% multiple compression scenarios in private equity and the credit markets

Who is Safe?

Things are moving so fast, it is hard to have high conviction on winners and losers in software. Higher value capture by agents, a transition to SaS and some degree of build over buy within entreprise undoubtably presents headwinds to seat-based subscription models. Hyperscalers have some protection from challenges to the seat-based subscription model, as for every seat lost to an agent, they can rent an extra GPU via their cloud (though by doing this they transition from an asset-lite to capital-heavy business).

It seems safe to exclude companies with a lot of IP and distribution, whether proprietary or licensed, like Netflix and Spotify. But then what if generative AI undermines the value of content libraries? There are similar concerns hitting videogame companies, which have the additional threat of world models, like Google's Genie 3.

Cybersecurity companies would seem to be winners, as a digital world populated by billions of agents will require all kinds of new protective measures. We have already seen examples of coding agents autonomously trawling GitHub repositories to surface long-overlooked vulnerabilities, but equally plenty of examples of security breaches relating to overly laissez-faire agent permissioning.

There is also a smaller cohort of companies whose data, software and hardware layers are tightly intertwined into products or platforms, and are therefore resistant to unbundling/disaggregation by agents. We highlight Planet Labs and Axon as two Horizon holdings sharing these qualities. 

In entreprise software, there will be stiff competition from a new breed of AI-native companies, not least being the frontier labs themselves. So much rests on the incumbents' ability to leverage their strengths in distribution and their willingness to innovate, even at the risk of cannibalising their existing revenue base. Against all expectation, Google appears to have succeeded in this so far, taking risks with their Search golden goose with AI overviews and AI mode that would have been unthinkable a few years ago. The value of support and accountability also shouldn't be underestimated in the many regulated industries of knowledge work, where the cost of being wrong is high. Finance, law, healthcare, public services - all of these industries will tolerate a degree of vendor lock-in in exchange for reliability and peace of mind. 

Perhaps the best hope for incumbent software companies more broadly is entreprise inertia, which may be sufficiently weighty to give them time to pivot into a more agentic SaS model. We are struck by how behind the broad perception of AI capabilities is outside of the close followers at the leading edge. OpenAI recently wrote a paper on this 'capability overhang', writing: "Despite paying for the same Plus plan, the typical power user (95th percentile) uses 7x more thinking capabilities than the typical user (median). This gap also persists at the message level, with power users requesting almost 3x more thinking capabilities per message than the typical user. This is a large gap and reflects the difference in how most people currently use ChatGPT. However, this measure understates the true size of the capability overhang. For example, the average OpenAI employee uses 15x more thinking capabilities per user and more than 5x more thinking capabilities per message than the typical Plus user".

These various uncertainties will resolve over time, and the outcomes will ultimately appear obvious with hindsight. Our main conviction is that agentic workflows will drive huge demand for inference, hence our focus on AI infrastructure.