Porter's Five Forces

Define the industry boundary

The industry is not "semiconductors" and not "AI". It is merchant accelerators and accelerator systems sold to hyperscale and enterprise data centres for AI training and inference, worldwide. That excludes consumer and edge GPUs, excludes CPUs, and crucially excludes captive in-house silicon such as Google's TPU or Amazon's Trainium, because those are never sold to a third party. Getting this line right is what makes the rest of the analysis work: AMD's MI-series is a rival because its price moves Nvidia's negotiation with a hyperscaler, while the TPU is a substitute and a backward-integration threat that constrains the merchant market from outside it.

Bargaining power of suppliers

Structurally very high. Leading-edge logic comes from essentially one foundry, TSMC, at roughly 90 percent of the advanced-node market; the true bottleneck for a year or more was CoWoS advanced packaging, not the wafer itself; HBM memory comes from only three firms with SK Hynix effectively setting the pace; and one level further up, ASML is a literal monopoly on EUV. None of these can be switched inside a tape-out and qualification cycle of twelve to eighteen months. The nuance a good candidate adds: the power is partly mutual, because Nvidia is the anchor customer on that node and prepays for capacity, so the supply chain takes only about a quarter of the end price (COGS is roughly 25 percent of revenue). High supplier power that is not currently extracting the pool is still a risk, because it can be exercised the moment the buyer stops being irreplaceable.

Bargaining power of buyers - the binding force

This is where the analysis lands. The buyers are four to six hyperscalers who are the most concentrated, best-informed and best-capitalised customers in commercial history, publish their capex budgets in advance, and each fund a credible backward-integration programme (TPU, Trainium and Inferentia, MTIA, Maia). Porter's checklist for high buyer power reads as a description of this customer set: few buyers, large share of the seller's revenue, sophisticated, and a real threat to make rather than buy. What holds them off today is switching cost, but note precisely what kind of switching cost it is: CUDA, the kernels, libraries and the engineers trained on them. A software switching cost is real, and it is also exactly the kind a buyer with 100,000 engineers and a $70bn capex line can afford to buy its way out of.

Substitutes and new entrants

The substitute is not another GPU; it is doing the same work a different way. That means custom ASICs designed with Broadcom or Marvell, and, quietly more dangerous, algorithmic efficiency: mixture-of-experts routing, distillation and quantisation all cut the FLOPs needed per unit of useful output, which shrinks the market without anyone entering it. On entrants, the classic barriers are weak here, because TSMC will sell capacity to any credible designer and the customers themselves will fund the development. The genuine barrier has moved from the chip to the rack: NVLink, the switch fabric, the reference system design and the software stack. State it that way and the strategic implication writes itself.

Competitive rivalry

Counter-intuitively low, and that is the entire reason for the 70-plus percent margin. There is no discounting war because supply has been short: buyers are allocated product rather than sold to, and a scarce good does not need a price cut. But scarcity is a cycle, not a structure. As CoWoS and HBM capacity expand, the competitive question flips from "who can get chips" to "who will cut price to fill a fab reservation", and gross margin is the first thing tested. Any Five Forces verdict here that does not separate cyclical scarcity from durable structure is wrong.

Where the profit leaks, and how the structure is moving

Walk $100 of accelerator revenue: roughly $25 goes to the supply chain (wafers, packaging, HBM, substrate, assembly), leaving about $75 gross; R&D and SG&A take perhaps $12 to $15; operating margin sits somewhere in the high 50s to low 60s (all approximate). Look one layer down and the picture is stranger: the hyperscaler earns a normal cloud margin renting that hardware out, and most of the model companies below them lose money. So the pool sits overwhelmingly at the chip layer, sandwiched between a monopolistic supplier that is not taking it and enormously powerful customers who are funding it. On the time dimension, three trends all point the same way: packaging and HBM capacity is expanding (rivalry rises), in-house silicon is shipping in volume (buyer power rises), and the workload mix is tilting from training to inference, which is more price-sensitive and far more ASIC-friendly.

Verdict and strategic response

Do not average the forces. Supplier power is high but restrained, rivalry is low but only because of scarcity; the binding force is buyer concentration paired with a credible backward-integration threat, and the two things holding it off (a software switching cost and a supply shortage) both have visible expiry dates. The structural response, which Nvidia is visibly running, is to change the unit of sale from a chip to a full rack and network, so that switching means re-architecting a data centre rather than swapping a component, and to widen the customer base into sovereign AI, neoclouds and enterprises so no buyer is 15 percent of revenue. For the investment committee, the underwriting question is not "can they build a faster GPU" but "does the system and networking lock-in compound faster than the four largest customers can port their workloads".

Define the industry boundary

The industry is routine, out-of-pocket-paid pathology (CBC, lipid, thyroid, HbA1c, vitamin D, liver and kidney panels) sold to walk-in and home-collection consumers in tier-1 and tier-2 cities of Karnataka, Tamil Nadu and the Telugu states. Deliberately excluded: radiology, which has a completely different asset base and capital barrier; hospital-captive labs, whose volume never reaches the open market and which are therefore a substitute rather than a rival; and esoteric or genomic testing, which is a separate industry where only three or four labs nationally can bid. The boundary test works cleanly here - when a funded entrant drops a 60-test package to Rs 999 in Bengaluru, this chain's walk-in volume visibly moves within weeks, so they are in the same industry.

Threat of new entrants

This is the force most students score as medium and it is in fact the one that decides the case. A collection centre is a 200 to 300 square foot franchised shop with a fridge, a centrifuge and a phlebotomist, costing roughly Rs 4 to 8 lakh to open (illustrative), and there is no meaningful licence gate beyond Clinical Establishments registration in most states. NABL accreditation is the only quality signal and it is effectively voluntary, because almost no consumer asks for it. A challenger does not even need a lab: third-party hubs sell processing capacity, so a new brand can be live in a city in about 90 days with rented science and a marketing budget. The practical consequence is brutal - the moment a pincode looks profitable, three more collection centres open on the same road inside a year.

Bargaining power of suppliers

Higher than the sector's reputation suggests, and it sets a hard cost floor. Analysers come from Roche, Abbott, Siemens Healthineers and Beckman on a reagent-rental model: the instrument is placed cheaply against a committed annual reagent purchase, and the systems are closed, so a Roche analyser runs only Roche reagents. That is a textbook razor-and-blade lock-in with a contractual minimum attached, and moving platforms means re-validating every assay and retraining the technicians. The second supplier group is MD Pathology doctors, who are legally required to sign reports and are in genuine short supply, which inflates salaries and caps how fast you can open hubs. Net effect: roughly 15 to 18 percent of revenue is locked into reagents and consumables before any competitive decision is made, plus a fixed pathologist cost per hub.

Bargaining power of buyers

The mistake here is to score the patient. Individually the patient is weak, but the market is fully price-transparent through aggregator apps, and for a commodity CBC there is zero loyalty. The buyers with real power are the intermediaries: the referring doctor or clinic that directs a large share of volume and expects service or economics in return; corporate wellness and insurer or TPA tenders, which are single large contracts awarded almost purely on price and routinely bid at 50 to 70 percent off list; and the health aggregator apps, which take roughly 20 percent of order value for handing over a customer who was never yours. So buyer power is high through concentration at the intermediary layer even though the end payer is completely fragmented - exactly the pattern Porter warns about.

Threat of substitutes

Three substitutes, none of which compete on price, which makes them worse. Hospital in-house labs bundle the test into the consultation, so the patient never enters this market at all. Point-of-care and home devices are quietly removing the most valuable volume: glucometers, point-of-care HbA1c and now continuous glucose monitors erode the repeat diabetic panel, which is the highest-frequency, highest-margin recurring test in the book. Third, insurers and employers increasingly fund preventive packages and route the whole cohort to one contracted partner, converting thousands of individual decisions into one tender you either win or lose entirely.

Rivalry, and the profit-pool walk

Rivalry is severe because the cost structure forces it. A hub lab is a fixed-cost machine - analysers, pathologist, night shift, sample logistics - while the marginal cost of one additional sample is only the reagent, perhaps Rs 40 to 80. When marginal cost sits that far below average cost, every player rationally discounts to fill capacity, and an entrant that does not need EBITDA this year gets to set the market price for everyone who does. Walk the rupee on an illustrative Rs 1,000 bill: roughly Rs 200 is discounted away to stay competitive, so Rs 800 is realised; the collection-centre franchise takes about Rs 240, or an app takes about Rs 160 if the order came through it; reagents and consumables are about Rs 130; phlebotomy and logistics about Rs 60; allocated hub fixed cost about Rs 150; marketing about Rs 60. What is left is roughly Rs 160 on a dense mature route and close to nothing on any hub running below about 500 samples a day.

Verdict and strategic response

Four forces are hostile and they are not equal: the binding pair is new entrants plus rivalry, because near-zero entry barriers combined with a high-fixed-cost hub is the precise recipe for a permanent price war, and buyer power through doctors, tenders and apps amplifies both. That means the 18 percent EBITDA is not a brand premium, it is a catchment-density outcome, and it will not survive a thin national roll-up. Three structural escapes are worth pricing into the deal: buy the number one or two position in three or four specific cities rather than a scattered footprint, because the economics are per-hub, not per-logo; shift mix toward specialty and esoteric tests where the supplier lock-in and know-how that hurt you now become your entry barrier and where three labs rather than three hundred can bid; and lock the referring-doctor channel with report and EMR integration, which is the only genuine switching cost this industry offers. Recommendation to the committee: pass at the asking multiple on a growth story, and re-price it as a two-state density consolidation with an esoteric bolt-on, or walk.