After years of enthusiasm surrounding artificial intelligence, investors are beginning to question the economics behind its unprecedented investment boom. Uncertainty around interest rates, rising financing costs and growing scrutiny over the returns on massive infrastructure spending have triggered a correction across the AI ecosystem.

Yet, while these concerns warrant attention, the underlying drivers of AI adoption remain compelling. As demand for computing power continues to expand and AI applications move closer to meaningful monetisation, the recent sell-off could be creating opportunities across both sides of the AI value chain: the infrastructure enabling its development and the software bringing it into everyday use.

When AI's spending boom meets the bond market

The AI investment boom continues to accelerate, with combined 2026 capital expenditure plans from Amazon, Microsoft, Alphabet, Meta, SpaceX and Oracle expected to go over $800 billion. However, uncertainty around interest rates and growing scrutiny over returns on investment have raised questions about the sustainability of this spending, particularly as the infrastructure buildout becomes increasingly reliant on debt financing. According to Bloomberg, AI-related borrowing reached approximately $460 billion so far this year.

These concerns came to the forefront in mid-October, when CoreWeave emerged as an early flashpoint for the broader AI sell-off. The debt-heavy cloud infrastructure provider saw the cost of insuring its debt against default surge above 1,100 basis points. Its difficulties have raised broader concerns around the neocloud business model, where substantial upfront investment, high leverage and reliance on external financing leave companies particularly exposed to rising borrowing costs and uncertainty over future returns. These credit-market fears spilled into equities, contributing to a sharp repricing of CoreWeave and other AI infrastructure companies.

However, neoclouds account for only a small fraction of overall AI capital expenditure, which is overwhelmingly driven by large technology companies with predominantly investment-grade credit ratings, diversified revenue streams and substantial cash generation. While credit concerns have also spread to these companies, the differences remain significant. Five-year credit default swap (CDS) spreads recently exceeded 300 basis points for Oracle, 200 for SpaceX and 100 for Meta, while Alphabet, Amazon and Microsoft remained below 100 basis points¹. Despite this widening, the stronger balance sheets and greater financial flexibility of the largest hyperscalers suggest that the broader AI buildout is considerably more resilient than the stress among highly leveraged neoclouds might imply.

The correction highlights a growing distinction between the strength of underlying AI demand and the financial risks associated with funding its expansion. While highly leveraged infrastructure providers have been particularly exposed, the broader valuation reset raises the question of whether markets are pricing in weakening fundamentals or simply becoming more selective about the economics of AI investment.

AI buildout: No inflection in sight

The biggest buyers are not slowing down. Alphabet, Amazon, Microsoft and Meta have guided to more than $700bn of capex this year, and Meta lifted the low end of its range in October. Spending on that scale lasts only if it earns its keep. Morgan Stanley estimates the return on invested capital in generative AI at roughly 25–46%, depending on how capacity is sold. At those returns, profits can fund a large share of the build-out.

Figure 1: Hyperscaler capex and estimated returns

Source: Company filings, as of 29/10/2026 and Morgan Stanley’s estimates. You cannot invest directly in an index. Historical performance is not an indication of future results.

Rental prices point the same way. Although token prices for some models have fallen, GPU rental rates have held up, and even the A100, launched in 2020, has become more expensive to rent this year. And in the recent earnings call, Microsoft still says demand of compute outruns supply, and the bottlenecks have not eased. Buyers who cannot get enough GPUs or memory accept higher prices and longer contracts. Recent earnings also suggested the constraints. Nvidia’s data centre revenue rose 117% in the quarter to July. SK hynix’s Q3 quarter revenue more than quadrupled, and it sees memory shortages lasting through 2030. Micron and the wider supply chain show the same strength. Further upstream, TSMC points the same way. Its third-quarter revenue grew 41% in US dollars, and 2nm chips went from 3% to 8% of wafer revenue in one quarter.

Demand and returns are holding up, yet the correction has left parts of the chain cheaper as earnings rise. That gap could offer a better entry point.

Figure 2: Token price decreased while the GPU rental prices were resilient

Source: Bloomberg, as of 29/10/2026. You cannot invest directly in an index. Historical performance is not an indication of future results.

Software strikes back: The rise of agentic AI

Early this year, investors priced software as AI’s first casualty. Most software is sold per seat, so an agent doing several employees’ work means fewer seats, slower growth and thinner margins. Some feared outright replacement, and Anthropic’s Cowork plugins deepened the sell-off. The BVP Nasdaq Emerging Cloud Index, made up mostly of software companies, fell 31.1% from the start of the year to its 10 April low.

Software has since proved harder to replace than feared. It holds the workflows, controls and proprietary data a company runs on, and an agent cannot easily copy them. Software companies are also building AI into their own products. From its April low to 29 October, the index rose 87.9%.

Figure 3: The cloud index outperformed the AI infrastructure index since it low in April (10/04/2026-29/10/2026)

Source: Bloomberg, as of 29/10/2026. You cannot invest directly in an index. Historical performance is not an indication of future results.

The recovery has been uneven, and stock selection matters more than sector exposure. Cybersecurity has led, helped by two AI trends. Attackers use AI to strike faster, and agentic AI adoption creates new vulnerabilities, since every agent is another identity to compromise. Both could keep cybersecurity spending resilient and reinforce its long-term investment case. Identity specialists such as Okta, managing access for people and agents alike, has returned 140.8% year-to-date at the end of October. CrowdStrike and Palo Alto Networks, the largest cyber pure players, has also performed strongly, and each have more than doubled year-to-date.

Usage-based pricing could offer an alternative route. Some software companies have reported their revenue grew with AI use, helped by consumption pricing, which bills an agent’s work rather than seats. Salesforce charges for Agentforce per conversation and through consumption credits alongside seats, and Agentforce’s annual recurring revenue passed $1.5bn in the quarter to July, up more than 240% year on year. Infrastructure software may also benefit, as AI platforms need tools to store data and monitor performance. DigitalOcean, for example, has more than doubled this year. But more broadly software has been mixed. For example, BlackLine and Klaviyo have each fallen by more than a third this year.

Going forward, personal agents could widen the market beyond the enterprise. The open-source agent OpenClaw showed the appetite in January. Meta’s Muse, launched in September, has taken the idea to scale. It books travel, shops and runs small business tasks, passed 10m users within about six weeks, and nearly a third of early users have linked a business account². More adoption of such personal AI agents can also add to infrastructure demand.

Anthropic's IPO: One catalyst, two AI opportunities

Anthropic's upcoming IPO could provide another catalyst for AI sentiment, offering investors a clearer picture of the commercial economics behind the technology. Its recently published S-1 filing revealed revenue of $16.3 billion in the first half of 2026 (Figure 4), up 1,300% year on year, with approximately 250 customers each generating more than $10 million on a trailing twelve-month basis. This rapid growth suggests that enterprise AI adoption is moving beyond experimentation towards large-scale commercial deployment.

The implications extend across both sides of the AI value chain. For infrastructure, Anthropic's compute and infrastructure expenses reached $8.9 billion in the first half, highlighting the enormous computing requirements needed to support growing demand. Yet these costs fell from 143% of revenue in Q2 2025 to 48% in Q2 2026 (Figure 4 a), suggesting improving operating efficiency as the business scales. Its extensive long-term compute commitments further reinforce the outlook for AI infrastructure demand, with agreements and planned spending exceeding $570 billion across major technology companies, including Amazon, Google, Microsoft, SpaceX, Broadcom and AMD, compared with approximately $86 billion across neocloud providers Nscale and Lambda (Figure 4 b). This concentration of spending among larger, more established technology companies provides greater visibility into future infrastructure demand, even as financing risks remain a concern.

Figure 4. Selected information from Anthropic’s Form S-1, filed on 26 October 2026

Revenue per dollar of compute

Source: Anthropic Form S-1, filed on 26 October 2026.

Overview of Anthropic’s Infrastructure Service Arrangements and Infrastructure Equipment Leases

Source: Anthropic Form S-1, filed on 26 October 2026; WisdomTree summary.

For software, Anthropic's growth offers further evidence that businesses are increasingly willing to pay for AI capabilities. As models become embedded in enterprise workflows and agentic applications take on more complex tasks, software companies could benefit from new consumption-based revenue streams and deeper customer engagement. Anthropic's IPO could therefore reinforce confidence in both the infrastructure powering AI and the software turning that computing capacity into commercial value.

Capturing both sides of the AI trade with WisdomTree’s AI-focused strategies

The recent market correction presents two potential opportunities for investors: broad exposure across both AI infrastructure and software, capturing the continued buildout and growing monetisation of AI, and more targeted exposure to AI infrastructure, where the valuation reset could offer an attractive entry point as underlying demand remains resilient.

The WisdomTree Artificial Intelligence strategy provides diversified exposure across the AI value chain, spanning semiconductors, computing infrastructure and software applications. While infrastructure-related holdings have faced valuation pressure, recovering software businesses have provided some resilience. Developed in collaboration with technology experts Nasdaq and the Consumer Technology Association (CTA), its underlying index leverages CTA's specialist expertise to identify and select companies based on their AI exposure and positioning within the value chain. This broad approach positions WTAI to capture opportunities from both continued infrastructure investment and accelerating AI software adoption.

For investors seeking more targeted exposure to a potential infrastructure recovery, the WisdomTree AI Infrastructure strategy offers a complementary approach. Developed in collaboration with SemiAnalysis, the strategy targets companies across the AI infrastructure supply chain, including semiconductor manufacturers, memory providers, networking specialists, advanced packaging and data centre and power infrastructure. With valuations having corrected despite resilient underlying demand, the strategy offers focused exposure to a potential rebound in AI infrastructure equities.

¹ As reported by Bloomberg based on data from ICE Data Services
² Axios: Meta courts small business with new Muse agent