Baron Global Opportunity Strategy | Q2 2026

We had an excellent quarter!

Baron Global Opportunity Strategy (the Strategy) gained 26.46% during the second quarter, compared to the 14.93% gain for the MSCI ACWI Index (the Index), and the 19.78% gain for the MSCI ACWI Growth Index, the Strategy’s benchmarks.

Year to date, the Strategy is up 18.42% compared to gains of 11.25% and 10.59% for the benchmarks, respectively.

Halfway through 2026, we observed some interesting similarities to last year in market patterns and the resultant outcomes. Entering 2025, we were coming off of two consecutive years of 25%-plus gains with a healthy/growing economy and tailwinds from investments and promises of AI. The geopolitical risks were elevated (something we thought was likely to continue for the foreseeable future) and the probability of a “breather” or a pullback seemed high. We got the “tariff tantrum” in mid-February and the Strategy experienced an approximate 25% drawdown over the next seven weeks. Then… it’s not like the tariffs were revoked and all of the trade issues were solved, but the market suddenly stopped caring about it. It’s as if it was suddenly “decided” that the tariff uncertainty and geopolitical risks would be figured out and that it would not be material in the long run. The markets rallied hard posting double-digit gains for the second quarter, while the Strategy had its best absolute return period since early 2020. Entering 2026… we were coming off of three consecutive years of 25%-plus gains with a healthy/growing economy and stronger tailwinds from even more investments and more promises from the benefits of AI. The geopolitical risks continued to be elevated, something we believe the markets are learning to live with for the foreseeable future, and the probability of a “breather” or a pullback seemed exceedingly high. In late February U.S. “and partner forces” had begun airstrikes against Iran. The Strait of Hormuz was “closed,” the price of oil soared, treasury yields spiked, and the markets sold off. Then… there was a Memorandum of Understanding (MOU) and a relative lull in hostilities. It’s not like there was a peace treaty or a new regime in Iran or free flowing traffic in the Strait of Hormuz, but the price of oil pulled back and it was “decided” that it will be figured out and that it’s unlikely to be material in the long run. The markets rallied hard posting double-digit gains for the second quarter, while the Strategy had its best absolute return period since early 2020. Similar indeed. We are oversimplifying of course, and there were lots of moving parts both this year and last, most notably the benefit of SpaceX. So, while history most certainly did not repeat itself, it did feel to us like it rhymed.
 

SpaceX

We had been under an NDA with SpaceX, which has prevented us in the past from fully explaining our investment thesis on the company. Now that SpaceX is public, we will attempt to explain why we are so excited about its prospects and long-term opportunities. We have often described ourselves as disruptive change investors and how disruptive change has become so critical and pervasive that we evaluate every investment opportunity through this disruptive change lens. We believe SpaceX stands at the intersection of two of the most consequential secular growth trends of our time: AI and Space Economy. The company is the most likely (obvious?) beneficiary of multiple early-stage, large S-curves disruptions. SpaceX has built a set of durable, multifaceted competitive advantages that to us appear to be insurmountable and is led by one of the most accomplished entrepreneurs and operators in history whose culture continues to attract extraordinary engineering and product-building talent. From the standpoint of uniqueness and opportunity set – it has no rivals. This is THE company with N=1.

More than a decade ago, SpaceX solved a problem experts said was impossible: instead of discarding rockets after each flight, it made them reusable. Today, SpaceX can reuse a booster (the bottom portion of the rocket) up to 35 times, while the rest of the industry still discards boosters after a single use. As a result, SpaceX needs only 3 boosters for 100 launches, compared with 100 for the industry. This created significant cost, scale, reliability, and innovation advantages, supported by higher launch frequency and the data generated from landed rockets. SpaceX has built a platform business with launch at its core. As its cost and scale advantages compound, the company can bring mass to space at lower cost and higher volume than anyone else, with the gap likely to widen further as it transitions to Starship. Lower unit costs expand markets, as we have seen across other industries: Moore’s Law enabled massive growth in compute, mobile, and internet; the falling cost of genome sequencing, from millions of dollars to roughly $200 today, helped power growth in biotechnology and pharmaceuticals; and lower-cost, increasingly capable intelligence tokens have expanded the AI market. SpaceX has reduced launch costs from the industry standard of nearly $20,000 per kilogram to an order of magnitude lower with Falcon 9, while Starship targets another major reduction to below $100. Beyond the expanding ecosystem that relies on SpaceX for launch, the company’s vertically integrated connectivity business, Starlink, has further supported Falcon 9’s scale because most missions carry Starlink satellites.

Speaking of Starlink… despite already serving more than 12 million subscribers, we believe the connectivity opportunity remains quite significant. Starlink delivers faster, lower-latency connectivity that works virtually anywhere. In a market of largely undifferentiated connectivity services – where many regions still have either no service or poor, dial-up-quality service – Starlink is disrupting an approximately $1.5 trillion connectivity market. Any investor who has experienced a Starlink-enabled flight will appreciate the significance of this advantage. Could Starlink ultimately reach 20% market share? 50%? 80%? Each 10 percentage points of share represents roughly $150 billion of revenue opportunity, with very high incremental margins. This is a high fixed-cost business: once the satellite constellation is in orbit, the incremental costs are primarily user terminals, customer acquisition, and customer support, and SpaceX continues to improve efficiency across the value chain. Starlink also creates the potential to connect more than 3 billion people globally who remain unconnected today⁵. Competition faces a steep challenge: smaller constellations offer less bandwidth, resulting in inferior service, while competitors also lack SpaceX’s scale and reusable launch economics, making their product more expensive to deliver. In addition, we believe SpaceX is well positioned to become an important player in global defense through Starshield, its dedicated government constellation.

Over the past five years, we have written extensively about AI and the disruption and the opportunities that we believe it represents. Now, with NVIDIA valued at over $4.8 trillion and hyperscaler CapEx approaching $750 billion annually, we believe AI is just getting started. The first agentic use case to make AI genuinely useful at scale has been coding, and even that market is still nascent. Annualized recurring revenue (ARR) is growing rapidly, with Anthropic surpassing $47 billion in May⁶, OpenAI passing $33 billion⁷, and overall generative AI revenues reaching $175 billion⁸. The coding opportunity alone is substantial: there are approximately 45 million to 50 million developers globally⁹ and that population should expand meaningfully as AI lowers the barriers to software creation. The number of natural-language coders could ultimately be far larger than the number of developers proficient in Python, Java, or C++. According to Ramp, which observes actual enterprise spending across its customer base, a top 1% firm already spends roughly $90,000 per employee per year on AI, or about half the median developer salary. A top 10% firm spends approximately $7,300, while the median firm spends just $137 – and all of these figures are growing exponentially.¹⁰ This suggests that coding alone could represent a multi-trillion dollar opportunity. If, over time, the median firm spends what today’s early adopters spend, the total addressable market could reach $4 trillion to $5 trillion. Extending that framework across roughly one billion information workers points to a potential opportunity measured in the tens of trillions of dollars. The range of outcomes is wide and will depend on the value AI creates, but the direction is clear: the more capable and lower-cost intelligence tokens become, the steeper the adoption S-curve is likely to be.

We believe SpaceX is emerging as a meaningful participant across multiple layers of the AI stack. The most immediate opportunity sits in the compute layer, where the company has recently signed its first terrestrial data center hosting agreements with Anthropic and Google, together representing $26 billion of annualized revenue. Based on our estimates, these contracts imply revenue opportunities in the mid-$20 billion per gigawatt for Anthropic and the mid-$40 billion per gigawatt for Google. If SpaceX scales to 10GW over the next several years, this could represent a $200 billion to $400 billion opportunity with very high incremental margins, assuming demand for intelligence tokens continues to exceed supply. We believe that is likely given the structural constraints around advanced chip production and memory availability.

Can SpaceX become a leader in AI hosting? We believe it already has, even with only its initial agreements with Anthropic and Google. SpaceX’s core competency is executing complex physical operations at extraordinary scale: reasoning from first principles, eliminating bottlenecks, removing unnecessary complexity, optimizing, automating, and moving with unusual speed. The company has already demonstrated the ability to build data centers faster and at lower cost than others, including a 100,000-GPU coherent cluster in 122 days versus an industry norm closer to two years. In a market advancing this quickly, time – is money! The sooner a data center comes online, the sooner it can begin producing valuable tokens and generating returns.

Still within the compute layer, SpaceX has a significant longer-term opportunity to place data centers in orbit. Picture satellites rather than buildings – powered by unlimited and free solar energy. While meaningful technical work remains, we believe SpaceX’s experience with Starlink satellites and terrestrial data centers positions it well to pursue this opportunity. Over time, as building data centers on Earth becomes more challenging and expensive because of regulation, power and cooling constraints, land scarcity, and construction costs, orbital data centers could benefit from effectively unlimited power, unlimited space, economies of scale, and a simpler regulatory framework. Starship is the prerequisite and therefore must come first. As for the scale of the opportunity, Elon has targeted 100GW per year by the end of the decade.

Over time, we believe orbital data centers could become the lowest-cost and most scalable architecture for AI inference. Their cost per gigawatt will depend on Starship reusability and the ratio of AI satellite weight to AI compute capacity. Based on the latest available information, Gen 1 satellites are designed to generate 160KW of AI compute per satellite, with the goal of increasing that capacity over time.¹¹ Even at 160KW per satellite, each Starship Version 3 carrying 28 satellites would deliver 4.5MW of compute, requiring roughly 223 launches for 1GW. At SpaceX’s target of less than $100 per kilogram, a Starship launch would cost less than $10 million, implying total launch costs below $2.23 billion for 1GW of AI compute. By comparison, five-year total cost of ownership for terrestrial data centers is roughly $10 billion to $15 billion per gigawatt (excluding AI compute costs in both cases on the assumption they converge at scale). This implies that 1GW of AI satellites (roughly 6.25 thousand satellites) could cost approximately $8 billion to $13 billion and still reach breakeven versus terrestrial data centers. That comparison also excludes the fact that terrestrial data centers incur ongoing power and operating costs that orbital data centers would largely avoid because power and “land” are free in space. For a 1GW terrestrial data center, power alone can cost roughly $1 billion per year, depending on efficiency and location, and terrestrial construction costs are likely to rise over time. Over the past three years, SpaceX spent roughly $10 billion on connectivity CapEx while launching approximately 8,000 Starlink satellites, suggesting this cost envelope for AI satellites is plausible, particularly because AI satellites should be simpler to manufacture: they replace complex phased-array connectivity hardware with solar panels and radiators.

Moreover, as Starship reusability improves and AI satellite manufacturing scales, orbital data centers should become increasingly cost efficient. Reusability is critical not only to launch costs but also to annual deployment capacity, because it enables far higher launch cadence. At the same time, terrestrial data centers are likely to become more complex and expensive as they scale. For every 100GW added to a future constellation (SpaceX’s annual target) charging just $10 billion per gigawatt, below current market rates, would generate $1 trillion of incremental revenue, again with very high incremental margins.

Despite these advantages, we do not believe hosting is part of the endgame of SpaceX’s AI ambitions. Elon’s objective is clearly to extend much deeper into the AI stack – the model layer through Grok as a frontier model and consumer service, and the application layer through Cursor, beginning with agentic coding, and MacroHard for enterprise AI. We view terrestrial hosting as a bridge: it can lower the effective cost of building higher layers of the AI stack while helping seed the orbital data center business. If successful, orbital data centers could become the lowest-cost source of AI inference and the only viable way to build AI at truly massive scale. That, in turn, could create a new AI platform for SpaceX, analogous to the launch platform it has already built, benefiting both its vertically integrated businesses in models and applications by providing the cheapest tokens at unmatched scale, and the broader ecosystem through lower cost, far more scalable AI hosting.

In case all of this was not exciting enough… additional longer-term opportunities include TeraFab for semiconductor manufacturing, point-to-point transportation and logistics, in-space manufacturing, space mining, the lunar economy, Mars, and more. We are NOT assigning any value to these opportunities and are simply pointing out that the positive optionality imbedded in SpaceX is unlike anything we have seen before in our investment careers. This is what we mean when we say that SpaceX is N=1!

We continue to encourage investors to evaluate the Strategy’s performance over a long-term horizon.

Space Exploration Technologies Corp. develops and launches advanced rockets, satellites, and spacecraft, with the long-term goal of making humanity multi-planetary. Shares rose as the company successfully completed the largest initial public offering in history, raising more than $85 billion. The proceeds are expected to accelerate growth across massive addressable markets, such as connectivity, launch, terrestrial and space infrastructure, and AI. Fundamental momentum was further reinforced by landmark compute hosting deals totaling tens of billions of dollars annually, including agreements with market leaders Anthropic and Google. SpaceX also announced its acquisition of Cursor, a premier enterprise AI-powered coding platform. Integrating Cursor’s technology, talent, and customer base provides another strategic stepping stone into the vast opportunities within AI applications and agentic systems. Lastly, the company conducted a successful test flight of the latest version of Starship, demonstrating meaningful advancements in rapid and full reusability. We believe these developments support sustained long-term revenue and profit growth well beyond current levels.

Observability and cloud monitoring platform Datadog, Inc. contributed to performance with the stock up 120.7% during the quarter after reporting its strongest results in years. Revenue accelerated to 32% year-over-year growth, marking the fourth consecutive quarter of acceleration and meaningfully exceeding consensus estimates. Operating margins and free cash flow also surpassed expectations. Management raised full-year revenue guidance by roughly $240 million against a $50 million beat, with next-quarter guidance implying continued strength. Critically, the strength was broad-based: non-AI core customer revenue growth accelerated into the mid-20% range at scale, while Datadog's AI-native customer cohort diversified and expanded to 22 customers spending more than $1 million annually. Net revenue retention reaccelerated into the low-120% range, and unit economics remained best-in-class with 80% gross margins and high-20% free cash flow margins. We view observability as a structural beneficiary of the AI buildout through cloud migration, faster application deployment, and an increasingly complex technology stack. These broadening growth drivers reinforce our conviction in the long-term thesis.

Leading cloud-native cybersecurity platform CrowdStrike Holdings, Inc. contributed to performance during the quarter with the stock up 95.6% as the market increasingly recognized the company as one of the primary beneficiaries of enterprise AI adoption. As organizations increasingly adopt AI, cybersecurity has moved to the forefront of C-suite concerns, and CrowdStrike is well positioned to potentially benefit as a leading platform with unmatched data and a large installed base that helps it defend better against breaches. That standing was reinforced when CrowdStrike was named a launch partner in major AI-industry security collaborations, including Anthropic's Project Glasswing and OpenAI's Trusted Access for Cyber program, underscoring its role in addressing new security challenges created by AI. This is already translating into demand. CrowdStrike’s AI detection and response product is experiencing one of the fastest growth trajectories in the company’s history. Management also raised full-year guidance meaningfully, reflecting the strength of its AI-driven pipeline. Because CrowdStrike's sensors already sit where AI agents operate, extending protection is a natural motion for existing customers and a relatively efficient cross-sell for CrowdStrike. These advantages reinforce our conviction in the company's long-term opportunity.

GDS Holdings Limited is a leading developer and operator of high-performance data centers, operating in key cities across China and expanding rapidly across Asia. Despite solid operating results with 340MW of new bookings year-to-date (on track to exceed their annual 500MW target) and strong performance earlier in the year, the stock declined 25.1% in the second quarter. Several factors weighed on performance, including management's commentary around the components of full-year guidance (which included a pull forward of a payment from its international subsidiary DayOne), a material step-up in expected capital expenditures over the next few years to support the robust growth opportunity, and a slight delay in the expected timing of an inflection in reported growth (which is driven by the natural variability of move-in dates). We added to our position during the quarter, seeing evidence of the growing AI wave in China through significant bookings growth and believe the market underappreciates the value of GDS's minority stake in its international subsidiary, DayOne.

Shares of Wix.com Ltd., a cloud-based platform for building and managing websites and online businesses, declined 51.0% during the second quarter. The stock came under pressure as growth in the company's core business decelerated, with bookings and new user cohorts softening relative to prior quarters. While the newer AI-driven offering (Base 44) continued to grow, it was not enough to offset slowing momentum in the established core business, raising questions about the durability of the overall growth algorithm. We had valued Wix for its combination of durable subscription revenue, expanding free cash flow, and disciplined capital returns, but the deceleration in the core business undermined a central pillar of our thesis. We have therefore exited the position during the quarter, redeploying capital toward opportunities with clearer growth trajectories.

Shares of Nu Holdings Ltd. underperformed during the quarter, declining 7.1% following a weaker-than-expected 1Q26 earnings release. Provisions exceeded expectations by 37%, resulting in a 7% earnings miss and raising investor concerns about the near-term outlook for credit costs and profitability. The unexpected departure of the CFO also weighed on sentiment. Concerns were further compounded by management’s ambitions in the U.S., where the competitive landscape is materially more mature and the long-term returns on incremental investment remain less certain. That said, the CEO subsequently outlined a more measured approach to U.S. expansion (limiting investment to 1% of revenues over 2026-2027) and capital deployment, which provided us with greater comfort regarding execution and resource allocation. We viewed the market’s reaction as an attractive opportunity to increase our position, taking advantage of the pullback in the share price. While these developments have tempered near-term expectations, we remain confident in Nu’s ability to capture a significant portion of the financial services industry across Latin America and beyond, supported by its enduring advantages in technology, data, cost efficiency, and customer experience.

The portfolio is constructed on a bottom-up basis with the quality of ideas and conviction level having the most significant roles in determining the size of each individual investment. Sector and country weights are an outcome of the stock selection process and are not meant to indicate a positive or a negative “view.”

As of June 30, 2026, the top 10 positions represented 57.9% of the Strategy’s net assets.

Our investments in the IT, Communication Services and Consumer Discretionary sectors represented 80.0% of the Strategy’s net assets, while Financials, Industrials, and Health Care sectors, as classified by GICS, represented another 17.8%. Cash accounted for the remaining 2.2%. Our investments in non-U.S. companies represented 38.4% of net assets, and our investments in emerging markets and other non-developed countries (Argentina) totaled 22.9% of net assets.

* Individual weights may not sum to the displayed total due to rounding.

During the second quarter, we initiated five new positions: the leading search and AI company, Alphabet, one of the leading memory providers, Samsung Electronics, a fast-inferencing AI chip provider, Cerebras, a supplier of critical parts used in aerospace engines and industrial gas turbines, DPC Holdings, and an engine provider for power generation, INNIO.

We also took advantage of stock price volatility and significant inflows into the Strategy to add to many existing positions including Shopify, Nu, NVIDIA, Amazon, GDS, ASML, TSMC, MercadoLibre, Tesla, ServiceTitan, and others.

We have sold four stub positions during the quarter: PDD, Wix, Viking Therapeutics, and Fiverr. We also reduced two other small positions: Afya and InPost.

Alphabet Inc. is the parent of Google, whose products are used by billions of people every day and include Search, YouTube, Android, Chrome, Gmail, Maps, and the Gemini assistant. In total, 13 Google products each serve over 1 billion monthly users, and 5 surpass 3 billion. The company reports in two primary segments: Google Services, which earns most of its revenue from advertising, and Google Cloud, which sells infrastructure, platform, and application services to enterprises. Within Google Services, Google’s core Search and Advertising platform is one of the most attractive technology business models ever built – distribution is effectively free, advertisers compete against one another to bid up prices, and users themselves determine which advertisers earn placement.

After emphasizing the importance of adaptability in recent quarterly letters, we ate our own cooking – so to speak – and initiated a position that likely would not have met our criteria in the past: Samsung Electronics Co., Ltd., one of the world’s three leading memory suppliers.

Before investing, we had to answer several key questions: Why is this a big idea? Has memory become less commoditized or less cyclical? Why Samsung, and why now?

Our Samsung thesis rests on four key points:

1. AI, and particularly agentic AI, is creating a large demand S-curve that remains in its early stages.
2. Supply growth is structurally constrained.
3. The memory industry appears to be becoming less cyclical.
4. Samsung’s non-memory businesses, including foundry, appear to be undervalued and provide both optionality and a partial hedge against Taiwan-related supply-chain risk.

First, on demand: unlike training, AI inference throughput (the number of tokens generated per second) depends heavily on how quickly a system can retrieve data and model weights from memory and perform the calculations needed to infer the next token. Greater bandwidth enables more tokens, which in turn drives higher revenue and better returns for AI factories. This dynamic is even more pronounced in agentic AI, where many internal “reasoning” tokens are generated for each external token delivered to a user. Agentic workloads also require much longer context windows (Micron has noted that context windows have expanded 30-fold year over year!),¹² as well as frequent tool use, including search, files, and applications. These use cases require more memory and more storage, particularly NAND, as users increasingly expect AI agents to remember prior interactions rather than restart from scratch. The result is the strongest memory demand environment in history, even before autonomous driving (autonomous vehicles expected to have 20 times the memory of non-autonomous)¹³, robotics, and other memory-intensive applications begin scaling in earnest.

Second, we believe memory supply will remain structurally constrained at least through the end of the decade.¹⁴ High-bandwidth memory (HBM), the stacked DRAM used alongside AI accelerators, is rapidly increasing as a share of total DRAM revenue, rising from roughly 8% in 2023 to about 33% in 2025, and should continue gaining share as models grow and agentic inference increases memory content per system. HBM is also significantly more wafer-intensive than standard DRAM, meaning more wafers are required per bit, which reduces total bit output from a given wafer capacity. At the same time, agentic AI is driving NAND demand, limiting the industry’s ability to convert NAND capacity into DRAM. New greenfield fabs take two to three years to build, complexity is rising, and capacity growth depends on constrained equipment availability, including a limited annual supply of extreme ultraviolet lithography (EUV) tools from ASML, which must also serve logic and foundry customers such as TSMC. Finally, incremental Chinese capacity is limited by restricted access to leading-edge equipment, including EUV, resulting in lower yields and lower quality.

Third, the memory industry may be becoming less cyclical. Memory suppliers are increasingly signing long-term agreements that provide better demand visibility and share volume and pricing risk with customers. In addition, memory – especially HBM – is becoming less commoditized as rising AI complexity makes co-design with accelerators more important. Co-optimizing the full system should increase differentiation and customer stickiness. Finally, as memory complexity rises, technology transitions to the next node are producing fewer incremental bits, making a larger portion of supply growth discretionary and giving the industry more flexibility to align bit supply with demand.

That brings us to the final question: why Samsung, and why now?

Samsung is the world’s largest memory manufacturer. In the first quarter of 2026, it led the DRAM market with a 38% share, ahead of SK hynix at 29%.¹⁵ After falling behind SK hynix during the HBM3E generation, Samsung recently regained momentum: in February 2026, it became the first company to begin commercial shipments of sixth-generation HBM4¹⁶ for NVIDIA’s next platform, and in March 2026, it signed an MOU with Advanced Micro Devices to supply HBM4 for Instinct MI455X GPUs alongside next-generation DDR5.

Samsung is also the only company with the full stack – DRAM, NAND, HBM, and a leading-edge logic foundry – under one vertically integrated roof, which should help it co-optimize memory and logic. Its internal smartphone and appliance businesses may also benefit on a relative basis because they can source Samsung memory at cost, reducing the impact of memory price increases compared with competitors.

We believe Samsung’s competitive advantages are durable. Memory is a three-player oligopoly with high barriers to entry, and Samsung and SK hynix together account for more than 80% of global HBM supply. Replicating the required 2.5D/3D advanced packaging, TSV stacking, and process leadership would take years and tens of billions of dollars, while the incumbents continue to advance. Beyond memory, we view Samsung’s display, mobile, automotive, and foundry businesses as meaningful optionality. Samsung Foundry is ramping its 2-nanometer node, with a 1.4-nanometer roadmap to follow. While the foundry business is currently loss-making, success with external customers could create meaningful profits and drive a re-rating. Samsung has historically traded at a conglomerate discount to memory pure plays, which we view as an inexpensive and underappreciated way to own a secularly growing memory business, with potential hedging value against global dependence on TSMC, HBM4 recovery as a near-term catalyst, and foundry as a long-duration call option.

The risks include mobile margin pressure, foundry execution, memory’s inherent cyclicality, and the possibility that AI demand does not meet expectations. Still, we believe the opportunity for durable AI-driven memory growth, combined with foundry optionality, offers an attractive risk-reward equation, particularly at around a 4x P/E multiple.

After following Cerebras Systems Inc. for years, meeting management several times, and touring its facility during our most recent Baron research trip to Silicon Valley, we participated in the company’s IPO during the quarter. Cerebras is a semiconductor company focused on fast AI inference – i.e. producing fast tokens. As inference evolves from one-shot responses to reasoning and agentic workflows, many internal tokens are generated for every external, customer-facing token. This shifts the primary bottleneck from raw compute, measured in FLOPs, to AI-factory throughput: the total number of tokens produced per second and the speed delivered to each user. Speed matters because latency compounds at every step of an agentic workflow; the faster tokens are generated, the more useful and attractive the AI agent becomes. As discussed earlier in the letter, we believe the AI inference S-curve remains very early. Founded in 2016 by Andrew Feldman, Cerebras is built specifically to address this fast-inference opportunity.

Cerebras’ Wafer-Scale Engine (WSE) is an AI accelerator the size of an entire silicon wafer, with substantial area dedicated to SRAM – the fastest available memory. The architecture is designed to keep as many calculations on-wafer as possible, going off-wafer only when necessary, which enables much higher bandwidth and as a result – faster token generation. The WSE-3 is 58 times larger than a leading GPU chip and, by keeping the model on-chip, can deliver inference up to 15 times faster than leading GPU-based solutions across various open-source models. Commercial validation includes a multi-year, 750MW agreement with OpenAI valued at more than $20 billion, with optionality for an additional 1.25GW; a partnership with Amazon to bring Cerebras’ fast inference to Amazon Web Services; and a second-quarter IPO that raised $6.4 billion. We believe Cerebras’ moat is durable because wafer-scale computing is difficult to replicate: the company has spent nearly a decade solving architecture, packaging, yield, and software-engineering challenges and remains the only company to commercialize this approach at scale. As agentic AI adoption grows and latency becomes increasingly important, Cerebras’ solution should become more valuable to customers.

We also initiated a new position in DPC Holdings Limited, a leading manufacturer of precision-cast superalloy components – the turbine blades, vanes, and structural castings that must perform under very harsh operating conditions including very high temperatures (1000c) and high pressure. These parts go into both commercial aircraft engines and industrial gas turbines (IGT) that generate electricity. The company occupies a hard-to-replicate market positioning as a critical supplier of the inner parts of an engine that only a handful of qualified manufacturers in the world are capable of producing.

Since the company began its turnaround in 2020 after over a decade under poor ownership, with the leadership of CEO Mike Quinn, who spent over a decade at Precision Castparts, and Co-Chairperson of the Board Dirkson Charles, who built and currently is CEO of aerospace and defense proprietary parts provider Loar (another holding of the Strategy), it has grown revenue at a high teens CAGR with expectations to continue that level of growth and even potentially exceed it. Mike brought the Precision Castparts toolkit, spent nearly $200 million retooling factories, replacing vacuum pumps, new wax presses, and established better contracts among other items to bring the company back to prominence.

With over 90% of EBITDA tied to the commercial aircraft and industrial gas turbine markets, DPC sits at the intersection of two large and durable demand s-curves. The first is the multi-year ramp in commercial aircraft production, as engine makers work through record backlogs that extend well past the end of the decade. The second is the rapid growth in the industrial gas turbine market as a result of extraordinary surge in electricity demand from AI data centers. Order backlogs at the major turbine original equipment manufacturers (OEMs) have reached record levels and now stretch several years out. These OEMs are now locking in capacity years in advance, driving long-term visibility with 70% of revenues driven by long-term agreements (LTAs). Most recently the company entered into 4 strategic customer partnerships – 2 with major IGT customers, 2 with A&D customers – worth an incremental $200 million of annual revenue. Additionally, customers are contributing up to 80% of capital investment dollars, leading to an attractive return on investment for DPC – will spend up to $150 million of CapEx to drive over $200 million of revenue with less than $75 million contributed by DPC.

Because its components are flight and safety critical, each part must pass through a multi-year qualification process tied to a specific engine program. The cost, time, and risk of qualifying an alternative supplier increases the switching costs and makes incumbent displacement exceedingly rare. Additionally, DPC’s vertical integration, across alloy, casting, coating, and machining and its ability to make 100% of its superalloys internally, reduces supply chain risks especially during the current tight supply environment. We expect revenue growth of high teens with EBITDA margins moving from the high teens to over 25% over the next five years as material latent operating leverage is unlocked and as the company expands its content in aerospace to blades and vanes, and continues to expand its capacity, serving the growing demand for power generation from AI datacenters.

As we are writing this letter, the hostilities in the Middle East have been renewed, the price of oil is rising again, and the AI “trade” has run out of steam and sprung a leak. It appears that more volatility is on tap, which has become par for the course. Despite the strong performance, we have not owned and do not own most of the really big winners of the last 12 months – the AI bottleneck names. We perceive many of them as lower quality businesses that historically lacked pricing power and were unable to earn their cost of capital throughfull market cycles. Perhaps it will be different this time and perhaps it will not, but we have chosen to remain judicious and uncompromising in refusing to reduce the quality hurdle in our idea generation even when significant short-term gains appear to be highly likely. As a result, perhaps it will improve our downside capture in times of inevitable market pullbacks, and perhaps it will not. We have no idea, and to be honest, we have never optimized for that.

Every day we live and invest in an uncertain world. Well-known conditions and widely anticipated events, such as Federal Reserve rate changes, ongoing trade disputes, government shutdowns, and the unpredictable behavior of important politicians the world over, are shrugged off by the financial markets one day and seem to drive them up or down the next. We often find it difficult to know why market participants do what they do over the short term. The constant challenges we face are real and serious, with clearly uncertain outcomes. History would suggest that most will prove passing or manageable. The business of capital allocation (or investing) is the business of taking risk, managing the uncertainty, and taking advantage of the long-term opportunities that those risks and uncertainties create.

We are optimistic about the long-term prospects of the companies in which we are invested and continue to search for new ideas and investment opportunities while remaining patient and investing only when we believe the target companies are trading at attractive prices relative to their intrinsic values.

Sincerely,