Separately, we have also exited SPDR Gold, the physical gold-backed exchange-traded fund (ETF) that has been a portfolio staple over the past year. The case for holding gold has become far less compelling even as geopolitical tensions continue to flare. Inflationary pressures from the ongoing Iran conflict have delayed expected interest rate cuts, removing a key pillar of support for gold prices that have already been stretched by speculative buying. Read our deep dive on gold, “What an elevated gold price is telling us and where to invest” in The Edge, April 6, 2026, for a comprehensive framework for valuing the metal.
Absolute Returns Portfolio — BUYS
With the proceeds from the above disposals, we have rotated into four new positions: Alphabet (Google), Microsoft, Nvidia and Talen Energy.
The Magnificent Seven names (Google, Microsoft and Nvidia) need no introduction. That said, it is worth discussing them individually, as they offer differentiated exposure to the secular tailwinds of growing AI adoption and other AI-led gains that are beginning to flow through to the bottom line of Corporate America. As key enablers of the AI ecosystem, these companies are among the primary beneficiaries of rising AI adoption.
See also: Investing where the economics take us: Today, it is America
Furthermore, each remains anchored by highly profitable legacy segments that continue to generate substantial and recurring cash flows. These core segments are not only being revitalised by AI but also provide the financial means necessary to aggressively fund and reinforce their next-generation AI investments, entrenching their sector leadership.
Alphabet (Google)
If history is any guide, it would be unwise to bet against Google’s capacity to innovate and stay relevant. After all, the current AI boom can conceivably be traced back to a Google research paper on machine learning entitled “Attention is All You Need”. Today, the company continues to embrace the Silicon Valley ethos of rejecting complacency, a philosophy embodied by its “Other Bets” segment, which aggregates the company’s “moonshot” ventures.
See also: To the stratosphere: The SpaceX IPO
To be clear, our investment case is not predicated on any single of these “moonshots” becoming the company’s next big S-curve. Rather, it is that Google offers unmatched exposure across layers of the AI value chain: hardware (via Google’s custom AI accelerator, the Tensor Processing Unit, or TPU), foundational models (such as Gemini, its flagship model family), cloud infrastructure (Google Cloud Platform), and a range of AI-enabled applications. These range from digital products such as NotebookLM, Google’s AI-powered research tool, to real-world deployments such as Waymo, its autonomous driving and robotaxi business.
Google’s AI ecosystem has continued to experience growing traction in recent quarters. On the infrastructure side, for instance, Google Cloud has consistently outgrown major cloud service provider (CSP) peers, as shown in Chart 1. Separately, the commercialisation of Google’s TPU business to external customers represents another emerging growth driver, with external sales expected to begin contributing to a modest but fast-growing share of revenue from the second half of 2026.
That said, it is important to keep these developments in perspective. Google remains predominantly an advertising-driven business (see Chart 2). Positively, and contrary to early fears of AI cannibalising traditional search revenue, evidence thus far points in the opposite direction. Google’s Search and Other Revenue expanded 19% year on year in the most recent quarter, driven by AI-enhanced experiences. Concurrently, AI-enabled ad campaigns have reportedly led to vastly improved targeting efficacy, supporting higher advertiser conversion rates.
Taken together, AI appears to be a clear tailwind for Google, benefiting both its core monetisation engine (advertising) and enabling the development of new growth drivers across adjacent end-markets. Against this backdrop, Google trades on a reasonable forward price earnings (PE) of 25, even as revenue and earnings per share (EPS) are expected to grow at 25% and 11% respectively this year (see Table 1).
For more stories about where money flows, click here for Capital Section
Microsoft
Of the Magnificent Seven, Microsoft has been the most punished by investor concerns over elevated spending. Over the past year, Microsoft’s share performance has remained lacklustre (see Chart 3, shown alongside our two other Magnificent 7 picks) even as a number of positive developments have emerged, creating an attractive risk-reward profile for entry.
Notably, Microsoft’s Intelligent Cloud segment has grown into a key growth engine for the company (see Chart 4), underpinned by demand for Azure cloud infrastructure and accelerating AI workloads across enterprise computing. The rapid growth of this segment, in turn, serves as a clear indicator of Microsoft’s progress in digital transformation and AI monetisation.
Separately, the renegotiated partnership with OpenAI has opened up new opportunities for the company. Under the revised terms, Microsoft will no longer pay a revenue share to OpenAI while retaining access to OpenAI’s intellectual property through 2032; meanwhile, OpenAI continues to share revenue with Microsoft through 2030. This structure frees up capital and cloud resources for Microsoft to prioritise its homegrown AI offerings.
Microsoft has also shown promise in its hardware strategy. The latest generation of its Maia AI chip reportedly delivers industry-leading performance for inference workloads versus comparable offerings from Amazon and Google. In early June, the company unveiled a new generation of consumer-facing AI PCs developed in collaboration with Nvidia, designed to enable on-device AI processing across laptops and desktops.
Looking ahead, Microsoft appears well positioned to benefit from the next phase of AI monetisation, supported by its extensive enterprise customer base and ability to cross-sell AI-enabled tools as usage continues to expand.
Nvidia
Last year marked a historic inflection point as Nvidia became the world’s first US$4 trillion and, in short order, US$5 trillion market-capitalisation company. At the time, the company appeared to balance on the knife’s edge of peak expectations, even as overlapping US and China export controls added further uncertainty.
Since then, Nvidia has continued to expand its addressable AI market via scaling its full-stack accelerating computing platform across core training and inference workloads, as well as emerging end-markets. Today, the company once again appears to offer a compelling entry point.
Over the past year, the company has moved from being primarily a graphics processing unit (GPU) supplier to actively shaping the contours of the AI infrastructure ecosystem. The company’s current roadmap is centred on an industrial vision of AI built around “AI factories” and the “AI grid” — concepts that position Nvidia as both the de facto systems architect and key infrastructure supplier of the growing AI economy.
Accordingly, this structural role positions Nvidia to continue capturing a disproportionate share of value as AI adoption broadens into enterprise, consumer and edge applications. The numbers tell the same story: in its latest April-ended quarter, top and bottom lines grew by 85% and 140% respectively.
Looking forward, the impending rollout of its next-generation Vera Rubin ecosystem that replaces the current Blackwell line provides a clear catalyst to sustain the company’s momentum. The architecture is purpose-built to support the rise of agentic AI, a theme that is currently gaining traction — we will expand on this shortly in the AI portfolio review below.
Talen Energy
As Nvidia’s CEO Jensen Huang has noted, AI can be thought of as a five-layer stack, with energy forming the foundational layer. Without power at the base, the layers above cannot function.
Talen Energy, an independent power producer (IPP) that operates primarily in the Pennsylvania-New Jersey-Maryland (PJM) region (one of the fastest-growing data centre hubs in the US), sits squarely within this framework and stands to benefit from rising electricity demand. The company operates 13.3GW of generation capacity across 12 facilities, with the Susquehanna nuclear plant as its crown jewel.
The Susquehanna operations currently represent Talen’s primary data centre exposure, supplying Amazon Web Services with stable, carbon-free baseload power under a long-term power purchase agreement (PPA). Moving forward, Talen is capitalising on growing data centre demand via a “hybrid” approach, combining power generation via its existing generation fleet with new-build projects. Management has outlined a pipeline of over 2GW of PJM interconnection submissions, alongside a land portfolio supporting an estimated 3GW-4GW of potential data centre capacity.
Crucially, more than half of Talen’s outstanding debt is fixed rate, providing meaningful insulation from interest-rate volatility. With IPPs having traded flat amid milder weather and regulatory uncertainty around grid interconnections recently (see Chart 5), we view this as an attractive entry point for a company that sits at a key bottleneck in the AI value chain.
AI Portfolio — SELLS
Moving on to the AI portfolio, our recent reshuffle is best understood by framing our holdings into three distinct buckets, each reflecting a key investment theme. Specifically:
1. Companies essential to the AI buildout: Amazon, Cadence Design, Hewlett Packard Enterprise, Marvell Technology and Datadog
2. Domestic substitution plays: Unusual Machines (US drone manufacturing), Alibaba (China AI stack) and Naura Technologies (China semiconductor equipment)
3. Companies reliant on external growth: Minth (auto parts) and Sieyuan Electric (grid equipment)
We have since exited the third category — Minth and Sieyuan Electric — in the light of growing geopolitical friction around Chinese firms expanding into overseas markets. Further, and consistent with our commentary on China’s hyper-industrialisation growth model, updated forecasts for Sieyuan Electric indicate sustained margin pressure, as competition intensifies within China’s domestic grid equipment supply chain.
Separately, we also briefly sold Marvell due to concerns of stretched expectations ahead of earnings, following a sharp run-up that saw the stock double in two monthss. But more on this shortly.
AI Portfolio — BUYS
Asking “What’s new in AI?” can seem a redundant exercise, as novel developments appear to crop up by the minute. Two structural themes have nonetheless begun to crystallise from the noise: agentic AI and edge AI deployments.
We have explored both topics previously, albeit in more conceptual terms. Briefly, agentic AI refers to systems capable of planning and executing complex, multi-step tasks with minimal human intervention, whereas edge AI refers to bringing inference closer to the end-user, which improves responsiveness and efficiency. Recent developments suggest both have begun moving rapidly towards monetisable deployment.
Accordingly, our four AI Portfolio additions lean heavily into these two trends, set against the backdrop of the ongoing AI infrastructure buildout — Akamai Technologies Inc, Marvell, Broadcom Inc and Roundhill Memory ETF.
Akamai
Historically, Akamai has operated as a content delivery network (CDN), hosting data on servers at the “edge” of the internet to enable faster and more reliable access to content. The company currently underpins 15%-30% of global internet traffic, supported by a network of more than 325,000 servers across 4100 points of presences (PoPs) and 131 countries — a footprint that vastly exceeds those of its closest CDN peers (see Table 2).
The importance of this physical footprint is central to the company’s current strategic pivot towards becoming a distributed cloud and edge computing provider. In 2025, the company launched the Akamai Inference Cloud to support AI workloads directly within its edge facilities. Already, this strategy is seeing growing traction, as highlighted by a landmark US$1.8 billion, seven-year deal with Anthropic for distributed compute, announced in tandem with first-quarter results this year.
Looking ahead, as AI deployment scales and workload complexity increases, demand for edge computing is likely to continue rising. In this scenario, Akamai is well positioned to benefit meaningfully given its established global footprint.
Marvell
Yes, you read that correctly: Marvell, in the space of a single week, was sold off, and then returned to the portfolio.
We have long mulled this question: how, and when, does one know it is time to sell? It is an increasingly pressing question in today’s crowded AI market, where the highest flyers often occupy an uncomfortable purgatory: feeling too rich
to hold, yet too consequential to sell.
Granted, “feeling” is the operative word here. As we said, US tech sector valuation does not necessarily appear unreasonable, provided the AI narrative and its attached growth expectations materialise (scan QR for more on this). That being said, it is certainly unsettling to watch stocks surge by 30%-50% in the span of days. As logic dictates, some degree of mean reversion is to be expected.
It was precisely this concern that led us to take profit on the stock ahead of its earnings release. Peers such as Arista Networks and Credo Technology had just suffered deep drawdowns despite solid earnings prints as elevated expectations left little margin for error.
But as the adage goes, a week in the AI world can feel like a decade elsewhere. While Marvell experienced a brief post-earnings wobble, developments over the remainder of the week have since reframed the narrative.
Building on an earnings report that highlighted accelerating data centre demand across all major product lines, Marvell’s Computex presentation showcased an expanded product portfolio, reinforcing its ambition to become a full-stack AI networking player.
This was further supported by Jensen Huang’s endorsement of the company at the same event, hinting at the potential for deeper partnership. Currently, Nvidia already has a US$2 billion investment in Marvell, alongside ongoing collaboration between the two that is focused on networking solutions for Nvidia’s AI factory and AI-RAN architecture (which targets edge applications).
Separately, Google’s announcement of a bumper US$80 billion equity raise, alongside reports of ongoing discussions with Marvell to develop new AI chips, has also contributed to the increasingly optimistic narrative surrounding Marvell.
The company’s data centre business is expected to grow by 50% or more year on year over the next two years, underpinned by strong optical networking demand and the ramp-up of the custom silicon programmes with Amazon and Microsoft.
Broadcom
Like Marvell, Broadcom is a fabless semiconductor company focused on custom-built chips. In addition, it also has a highly lucrative enterprise software segment anchored by VMWare, a virtualisation platform widely used across enterprise data centre and hybrid cloud environments.
In custom AI chips (or ASICs), Broadcom and Marvell collectively account for roughly 80%-90% market share. Between the two, Broadcom remains the undisputed leader with an approximately 70% share and a market cap of nearly US$2 trillion.
In its most recent quarter, Broadcom’s CEO Hock E Tan reaffirmed AI semiconductor revenue “in excess of US$100 billion in 2027”. While the markets were unimpressed by the absence of an upward revision to this forecast — triggering the stock’s post-earnings plunge — we remain optimistic on Broadcom’s longer-term prospects.
The company remains one of the most strategically entrenched providers of custom silicons that power AI workloads: Google, OpenAI, Anthropic and Meta are part of its roster of key partners. Looking ahead, while Nvidia GPUs are expected to retain the majority share of AI workloads, ASIC growth is expected to outpace GPUs over time, given the power and cost advantages they provide across narrow, highly-specialised workloads. Bloomberg Intelligence estimates that the ASIC market will continue growing at a 27% CAGR (compound annual growth rate) through 2033.
Separately, Broadcom continues to lead the AI networking market, and revenue in the segment is expected to triple this year, underpinned by an ongoing upgrade cycle from 800 Gbps to 1.6 Tbps optical speeds. Shipments of its Tomahawk 6 Ethernet switch chipset, designed to support this next-generation bandwidth, began in March this year, and is expected to gain traction through the year as hyperscaler deployment continues to scale.
Roundhill Memory ETF
Rounding up our AI Portfolio additions is the Roundhill Memory ETF. The AI memory trade has been on a tear in recent months, as high-bandwidth memory (HBM) has emerged as a bottleneck in AI infrastructure. By rough analogy, HBM acts as the “short-term memory” of an AI system. While GPUs perform calculations, HBMs ensure that they are continuously fed with data, thus maximising utilisation and performance.
The key question, of course, is whether the industry is genuinely running short of memory.
At the hardware level, HBM consists of multiple dynamic random access memory (DRAM) dies stacked vertically into a single compact module using advanced packaging techniques. At present, industry estimates point to DRAM markets remaining constrained through 2028 (see Chart 6), supporting elevated memory chip prices for the period.
That said, even as this bottleneck persists into the medium term, a number of emerging technologies are already vying to erode, or potentially replace, HBM’s dominance. As such, rather than attempting the fool’s errand of pre-emptively picking out a single winner, we have opted for diversified exposure across the memory sector via the Roundhill Memory ETF, which provides exposure across the major players of this space, most notably SK Hynix Inc, Samsung and Micron Technology.
