Tempus AI has received 510(k) clearance for ECG-PH, an algorithm that reads a resting 12-lead ECG and returns a binary flag for elevated mean pulmonary artery pressure above 20 mmHg. It is the third cleared device in the company’s cardiology line, after ECG-AF and ECG-Low EF, and it follows the same template: no new hardware, no new consumable, no new patient encounter, just a model running over a recording the health system already captured for other reasons. The label restricts it to symptomatic patients aged 40 and older with no known PH history, excludes paced rhythms, and forbids serial monitoring or standalone diagnosis.
The interesting number is not in the press release. It came three weeks earlier, on the July 30 earnings call, when the CFO was asked whether the roughly 40% algorithm attach rate on solid tumor had held. He said attach had risen to 45% in Q2, broad-based across the suite, and then added that many of those algorithms are not paid, so the number of algorithms ordered has no revenue impact and none is embedded in guidance. That is the company telling the market, unprompted, that its algorithm portfolio is currently a zero on the income statement.
This is not damning in oncology, because in oncology the attach model has a coherent economic logic. Tempus performs the sequencing. The order is billed. The algorithm rides along on a test the company already collects money for, costs almost nothing to serve, and makes the underlying assay harder to displace. A 45% attach rate is a measurement of switching cost, not of demand for software. The right way to read it is as a retention statistic that happens to be denominated in features.
Cardiology breaks that structure at the root. Tempus does not perform the ECG. The recording is captured on a GE, Philips or Hillrom cart, stored in a cardiology information system, and read inside Epic or Cerner. There is no Tempus-billed base event for the algorithm to attach to. Which means the three cleared ECG devices have to be sold as something the oncology algorithms never had to be: an actual product, with an actual price, through an enterprise software motion into hospital IT budgets, or through a reimbursement pathway that a 510(k) clearance does not confer, or as a pharma-funded case-finding program. Those are three different businesses with three different sales cycles, and the company has disclosed revenue for none of them.
The moat here is real but narrower than the headline suggests, and it is not the model. Architectures for ECG-derived structural heart inference have been in the literature since the Mayo work, and Anumana, AliveCor and Eko are all in the same water. What Tempus owns is the labeled corpus and the regulatory file. ECG-Low EF was trained on more than 930,000 ECGs across over 170,000 individuals, which is the kind of paired ECG-to-catheterization or ECG-to-echo dataset that is expensive to assemble and impossible to scrape. On top of that sits a compounding regulatory asset: each cleared device becomes a predicate, and the marginal cost and timeline of device four and device five fall accordingly. Three clearances in roughly two years is evidence the pipeline mechanism works. The vulnerability is at the other end of the stack, where GE and Philips own the acquisition device itself and can bundle inference into the cart at zero incremental price. Whoever owns the machine can always give away the software.
The epidemiology also deserves less credulity than it usually gets. PH affects around 1% of the population and up to 10% of adults over 65, but a binary classifier aimed at a 1% base rate produces a positive predictive value no clinician would act on. The label solves this without the model doing any of the work: by restricting use to patients over 40 who already present with dyspnea, fatigue, chest pain or edema and who have no known PH, the indication raises pre-test probability into a range where a binary output is usable. The enrichment lives in the intended-use statement, not in the algorithm.
Then there is the question of who captures the value of a positive result. The confirmatory pathway runs to echocardiography and then to right heart catheterization, and Tempus owns neither. Worse for the monetization story, mPAP above 20 mmHg is the broad hemodynamic definition, which sweeps in Group 2 pulmonary hypertension from left heart disease and Group 3 from lung disease. Those are the bulk of cases and the answer is to treat the underlying condition. The drug-funded slice, where sotatercept, the Tyvaso franchise and the Actelion oral portfolio compete for diagnosed patients, is Group 1 PAH plus PH-ILD, a small fraction of what a wide net catches. A pharma-sponsored screening deal is plausible and would be the cleanest path to a stated price per read. It would also be sized off a much smaller population than the 1% figure implies.
The stock is not trading on any of this. Shares closed Monday at $65.75 for a market capitalization near $11.9 billion, having traded as high as $73.50 intraday and given back better than ten percent from the session peak. The move that got them there had nothing to do with Tempus: on August 19 the stock rose 23.9% from a $49.36 close on Merck and Moderna’s Phase 3 readout for intismeran autogene in resected melanoma, which validated the sequencing platform inside Personalis, the company Tempus agreed in July to buy for roughly $1.5 billion in stock. Short interest running near 20% into that print did the rest. The stock sits about 37% below the October 2025 record of $104.32 and above the consensus target of $62.64, which sixteen analysts have been trimming rather than raising.
Against $1.595 to $1.605 billion of 2026 revenue guidance, $820.7 million of cash and $460 million of converts, the enterprise trades at roughly seven times forward sales. Personalis adds about 12% dilution for something near $90 million of annualized revenue, which raises the blended multiple even allowing for that asset’s 442% year-over-year clinical growth. Q2 itself was the strongest print the company has delivered: $382.5 million of revenue up 22%, 64.4% gross margin, a first GAAP profit of $5.6 million against a $42.8 million loss a year earlier, and roughly $200 million of bookings.
Base case is $58 to $72, the range the stock has occupied since the squeeze, with the Personalis close in late Q4 or early 2027 as the gating event and cardiology contributing nothing identifiable. Bull case is a retest of the $85 to $95 zone, which requires the Q3 print to show Data and Applications sustaining high-twenties growth and, separately, a named health system or pharma contract that puts a price on an ECG algorithm for the first time. Bear case is cohort derating to $42 to $48, the July low, and it does not require anything to go wrong at Tempus: seven times forward sales on a business turning its first $5.6 million of GAAP profit is a multiple held up by the same risk appetite currently supporting every other unprofitable AI name, and it compresses on rate or sentiment without a single missed number.
The disclosure that settles the cardiology question is narrow. At the next call the attach-rate question will be asked again, and the answer will either repeat that the algorithms are not paid or it will not. The first algorithm Tempus reports as separately reimbursed, with a fee and a volume, converts a clearance count into a business. Until then, three cleared devices is a regulatory achievement the income statement has never seen.