TheEngineRoom

NVIDIA designs the GPUs that power artificial intelligence — both the training of large language models and the inference that runs them in production. The H100 and B100/B200 Blackwell chips are the gold standard for AI compute, and the demand from hyperscalers (Microsoft, Google, Amazon, Meta) and sovereign AI programmes vastly exceeds supply. But the hardware story understates the moat. CUDA — NVIDIA's proprietary software ecosystem — is the reason developers can't simply switch to AMD or Intel when chips become available. CUDA has been taught in universities for 15 years, is embedded in every major AI framework, and represents billions of developer hours of optimisation. The GPU is the product; CUDA is the lock.

AMD has executed one of the most remarkable turnarounds in semiconductor history. Under Lisa Su, it went from near-bankruptcy in 2014 to dominating Intel in server CPUs with EPYC, and building a credible AI GPU business with the MI300X and MI350 series. The EPYC business is already a major profit engine — server CPUs are AMD's most valuable near-term compounder, with continued share gains from Intel's manufacturing stumbles. The AI GPU story is earlier but potentially transformational: the OpenAI partnership signals that hyperscalers are willing to invest in a second source for AI compute — reducing dependency on NVIDIA and giving AMD a real opportunity in the multi-hundred-billion-dollar AI accelerator market.

Broadcom is the infrastructure backbone of the AI revolution in two distinct ways. First, custom silicon: it designs bespoke AI accelerator chips (XPUs) for Google (TPUs), Meta, and other hyperscalers who want to avoid total dependence on NVIDIA. Each custom chip programme is a multi-billion-dollar long-term engagement with switching costs that are essentially infinite — Broadcom becomes structurally embedded in its customers' AI architectures. Second, Ethernet networking: Broadcom dominates the networking chips that connect the thousands of GPUs inside an AI data centre — as GPU clusters scale, network chip demand scales with them. The VMware acquisition adds $8bn+ of high-margin recurring software revenue that cushions semiconductor cyclicality and funds continued R&D investment.

Arista Networks builds the switching and routing infrastructure that connects cloud data centres — the plumbing inside hyperscaler facilities. After winning the first wave of cloud networking from Cisco in the 2010s, Arista is now the primary beneficiary of a second wave: AI cluster networking. Training a large AI model requires thousands of GPUs to communicate simultaneously at extremely low latency. This demands specialised, ultra-high-speed networking — Arista's 400G and 800G switches. Major hyperscalers (Microsoft, Meta, Google) have multi-year purchase commitments for Arista's AI networking products. The EOS software platform — a single network operating system running across all Arista hardware — creates meaningful switching costs and a growing software revenue stream on top of the hardware business.

TSMC is where chips are actually made. NVIDIA designs its GPUs; TSMC fabricates them. Apple designs its M-series chips; TSMC fabricates them. AMD, Qualcomm, Broadcom, MediaTek — the same story. TSMC is the world's largest and most advanced contract chipmaker, commanding approximately 90% of the market for the most advanced process nodes (3nm, 2nm). No other company on earth can manufacture leading-edge semiconductors at TSMC's yield, volume, or quality. This is not a commodity manufacturing business — it is a monopoly of technical capability accumulated over 35 years and hundreds of billions of dollars of R&D. Every AI chip that matters runs on TSMC silicon.

Alphabet is three businesses in one. Google Search: the world's dominant information retrieval system, generating ~$200bn+ in advertising revenue annually with structural advantages in data, distribution, and AI-powered relevance that are not easily disrupted. YouTube: the dominant video platform globally with growing subscription and advertising revenue. Google Cloud: now the world's third-largest cloud provider and growing at 28%+ with expanding profitability — catching up to AWS and Azure from a lower base. The AI angle is existential: Gemini is Alphabet's response to ChatGPT, and the AI Overview feature in Search monetises AI at scale through existing advertising infrastructure. The bear case on AI disrupting search has so far failed to materialise at the revenue level.

Amazon is a case study in reinventing revenue. AWS, which started as an internal tool, is now the world's largest cloud provider with $100bn+ annual revenue growing at 17%+ and expanding margins — it funds the entire enterprise. Advertising, almost invisible to consumers but worth $50bn+ annually, has become the third pillar — Amazon's unique position inside the purchase funnel makes its ad inventory uniquely valuable to brands. The e-commerce business is now optimising efficiency after years of heavy investment, releasing a wave of free cash flow. The AI angle is multi-layered: AWS sells AI compute (including NVIDIA GPUs and Amazon's own Trainium/Inferentia chips), offers Bedrock for AI model access, and is integrating AI into Alexa and the fulfilment network.

ServiceNow started as an IT service management platform — the system that handles internal IT tickets, change management, and incident response. It has since expanded to become the central workflow automation layer across HR, finance, legal, and customer service for the world's largest enterprises. Every Fortune 500 company is either a ServiceNow customer or a prospect. The platform's genius: once implemented, it becomes the connective tissue of an organisation's operations — integrating with hundreds of other software systems, becoming the single pane of glass for all enterprise workflow. AI is now accelerating the thesis: ServiceNow's Now Assist AI features are embedded directly in workflows, generating expansion revenue from existing customers without new sales cycles.

Eli Lilly discovered and commercialised tirzepatide (Mounjaro for diabetes, Zepbound for obesity) — the GLP-1/GIP dual agonist that has demonstrated superior weight loss efficacy versus both Novo Nordisk's semaglutide and all prior obesity treatments. The addressable market for obesity treatment alone is estimated at over 1 billion people globally — a genuinely transformational healthcare opportunity. LLY's edge over Novo Nordisk: superior clinical efficacy (tirzepatide consistently outperforms semaglutide on weight loss outcomes), a deeper pipeline (orforglipron, retatrutide — oral and next-generation GLP-1s), and faster manufacturing scale-up. Financial metrics are exceptional: exceptionally high gross margins (80%+) and operating margins (~45%), with revenue growing at roughly 50%+ in recent periods in recent periods. The patent runway extends well into the 2030s.

Intuitive Surgical has the most durable competitive moat in medical devices. The da Vinci robotic surgical system is installed in over 10,000 hospitals worldwide, has been used in nearly 17 million procedures, and has trained a global generation of surgeons who know no other robotic platform. The razor-and-blades model is exceptional: hospitals pay $1–2m for the system (razor), then spend $1,500–$3,500 per procedure on proprietary single-use instruments and accessories (blades) — recurring revenue that grows with every installed system and every additional procedure type approved. The da Vinci 5 upgrade cycle is underway, driving both system replacements and capability expansion. New platforms (Ion for lung biopsy, SP for single-port procedures) extend the addressable procedure market.

Merck is the maker of Keytruda — the world's best-selling cancer immunotherapy drug, approved across 40+ cancer indications and generating $25bn+ annually. Keytruda's US patents expire in 2028, threatening roughly 40–47% of Merck's current revenue. This is the dominant concern in the market, and it has compressed Merck's valuation to decade lows — currently around 12× forward earnings, historically cheap for a business of this quality. The investment case: Merck has invested $70bn+ since 2021 in pipeline diversification (Winrevair for pulmonary hypertension, islatravir for HIV, Welireg for kidney cancer), the subcutaneous Keytruda formulation extends patent protection significantly, and a new oncology pipeline should replace meaningful revenue by 2030.

Rolls-Royce makes the engines that power the world's largest widebody aircraft — the Airbus A350, Boeing 787, and A330. Its business model is distinctive: engines are sold at near-cost or at a loss, and the real money comes from Long-Term Service Agreements (LTSAs) — multi-decade maintenance and overhaul contracts priced by the hour flown. Every hour a Trent engine flies, Rolls-Royce earns. As post-COVID flying hours recovered to pre-pandemic levels and beyond, Rolls-Royce's LTSA revenues surged. CEO Tufan Erginbilgin executed a cost restructuring programme that simultaneously expanded margins — the combination has produced a company that in 2024/25 generated £3bn+ in free cash flow from a business that was burning cash 3 years prior. Defence is a structural tailwind: Rolls-Royce makes nuclear reactor systems for Royal Navy submarines, and the global defence spending revival is accelerating new contracts.