The Theranos Autopsy

3. Claimed versus true

4. What was real, and what still stands

A fair autopsy has to list what did work, because a new entrant should keep those parts.

• Some immunoassays did run on a small device. The Edison could perform a handful of immunoassays on small, diluted samples. The HSV-1 IgG test was good enough to pass FDA review in July 2015 using fingerstick samples in nanotainers. Small-volume immunoassay in a box was not the impossible part; narrow-menu devices from Abbott, Roche, Siemens and others had been doing it for years.
• The miniLab concept was sound. The 2016 design put sample handling, a centrifuge, optical detectors for chemistry and immunoassays, a cytometer and a nucleic-acid amplification module in one chassis, run by cloud software. That is almost exactly the architecture Truvian cleared with FDA in 2025-2026, that Genalyte's Merlin uses, and that biosurfit's spinit has sold in Europe since 2012. Theranos was not wrong about the shape of the box. It lied about how far along it was.
• The retail idea was right and has survived. Walgreens agreed in 2018 to host at least 600 LabCorp patient service centers. Quest draws blood inside Walmart stores. Kroger rolled fingerstick lipid-and-glucose analyzers out to its more than 2,000 pharmacies and clinics. Babson Diagnostics collects fingertip blood in Austin pharmacies with BD and runs 68 tests. The demand for lab testing in a pharmacy without a phlebotomist was real.
• Transparent cash pricing was right. Theranos published a price list at well below Medicare rates. Babson now posts a CBC at $25 and a metabolic panel at $19; Function Health sells 160-plus tests for $365 a year. Posted prices are now normal in consumer lab testing.
• The Arizona direct-access law outlived the company. Patients in Arizona can still order their own tests. Many states now allow some form of direct-access testing.
• The patents exist and are enforced. About 700 patents went to Fortress. Two were asserted against bioMérieux's BioFire unit in 2020. Any new entrant should do a freedom-to-operate review against that portfolio, not because the ideas worked at Theranos but because the claims are broad.

5. Why the science failed

5.1 Capillary blood is not venous blood

A fingerstick sample is capillary blood mixed with interstitial fluid and damaged cells, collected under pressure, in small volume, exposed to air. The literature is consistent:

• Hemolysis (ruptured red cells) occurs in about 17.7 percent of capillary samples versus 10.6 percent of venous samples; in heel pricks it can exceed 30 percent. Hemolysis releases potassium, lactate dehydrogenase and AST into the plasma.
• In a 2021 study of 59 paired fingerstick and venous samples from adult inpatients, results were comparable for most of a basic metabolic panel except potassium (higher in capillary) and bicarbonate (a 5.5 mmol/L negative bias from underfilled tubes). The authors concluded that pre-analytical problems "preclude widespread use" in that population.
• A 2024 comparison of 34 routine chemistry analytes found excellent agreement for most, but capillary samples hemolyzed significantly more often (157 vs 106 samples with a hemolysis index of 1 or more).
• Platelet counts run lower in capillary blood. Wet alcohol on the skin ruptures cells. Squeezing ("milking") dilutes the sample with tissue fluid. Even with good technique there is a large variation between successive drops.
• An upper-arm capillary collection device produced samples with consistently low hemolysis indices, and bias against venous blood mostly within a few percent (the full range across analytes was minus 4.4 to plus 14 percent). That is why the modern honest players collect from the arm or use standardized fingertip technique with trained staff.

Theranos's answer to all of this was dilution, which multiplies every error, and silence.

5.2 One drop cannot feed every assay class

Different tests need different sample preparation and different anticoagulants, and the tubes are not interchangeable:

The honest designs split the sample at collection into two or three fractions with the right additives and route each to the engine that needs it. Theranos tried to run everything from one undifferentiated nanotainer.

5.3 Mass balance

A drop of blood is roughly 20 to 50 microliters. Collecting enough for a lipid panel and a chemistry panel takes 16 to 18 drops (about one tenth of a venous draw), trained staff and a purpose-built collector such as BD's MiniDraw. Theranos marketed its nanotainer around "a few drops." A conventional immunoassay consumes 20 to 100 microliters per analyte; a chemistry disc wants about 100 microliters for a 14-analyte panel; a CBC needs 15 to 30 microliters. Two hundred tests from one drop violates arithmetic unless every assay is multiplexed onto one shared reaction (which is what label-free photonic chips and multiplex immunoassays attempt, with dozens of analytes, not hundreds) or miniaturized to nanoliter scale, where matrix effects and sensor noise dominate. Even the 2016 miniLab could not run eleven assays on 160 microliters.

5.4 Dilution and dead volume

Commercial analyzers have minimum fill volumes and dead volumes (the HORIBA Microsemi, for example, uses 18 µL for analysis but needs 60 µL in the tube). Theranos diluted fingerstick samples to reach those minimums on Siemens machines, which moved concentrations toward the bottom of each assay's validated range, where imprecision is worst, and amplified any pipetting or hemolysis error by the dilution factor.

6. Why the company failed

The science problems above are solvable with narrower menus, better collection and honest validation, which is what later companies did. Theranos failed as a company for governance reasons:

• Secrecy in place of peer review. By August 2016 a PubMed search for "Theranos" returned 14 papers, two with original data from company scientists. Every competitor with a real product publishes method comparisons.
• A board without laboratory medicine. Statesmen and generals, no clinical chemists, no pathologists.
• An unqualified lab leadership structure. Balwani, with no laboratory training, oversaw the lab. One credentialed director testified that he resigned over accuracy concerns; a dermatologist with no laboratory background later held the director title.
• Retaliation against the people who were right. Tyler Shultz and Erika Cheung were threatened with litigation for reporting proficiency-testing problems that CMS later confirmed.
• The LDT loophole used as a shield. Running tests only inside your own CLIA lab legally avoided FDA premarket review in 2013-2015, but Theranos used the loophole to avoid validation rather than to do it carefully.
• Partner deception. Walgreens and Safeway committed hundreds of millions on demos and claims that were staged.

7. What exactly was fraud

Not every failure was fraud, and the courts drew the line precisely.

• Investor fraud (convicted). Holmes was convicted on three counts of wire fraud and one of conspiracy for statements to investors about the device's capabilities, the Pfizer and Schering-Plough validations, the military deployment and the revenue figures. Balwani was convicted on all twelve counts, including the patient counts.
• Patient-facing claims (Holmes acquitted, but harm documented). The jury did not convict Holmes on counts about defrauding patients. CMS, separately, documented inaccurate results that reached patients and forced the voiding of two years of Edison results.
• Regulatory misfiling. The nanotainer was marketed as a Class I exempt device when FDA considered it Class II requiring clearance.
• Scientific misrepresentation. Pharma logos on validation reports the companies never wrote.
• Staged demonstrations. The error-suppressing demo mode and the practice of running visitor samples on commercial machines.

The lesson for a new company is that the line between failure and fraud is disclosure. A device that cannot yet do fingerstick chemistry is a development program. A device sold as doing it is fraud.

8. The money

• Raised: more than $700 million per the SEC; totals near $945 million across all rounds are often reported.
• Peak valuation: about $9 billion (2014).
• Walgreens: $140 million paid in; settlement under $30 million recovered; $44 million more paid by Walgreens to its own customers in 2023.
• Safeway: about $350 million in clinic build-outs that never hosted Theranos testing.
• Arizona refunds: $4.65 million.
• Restitution ordered against Holmes and Balwani: $452 million.
• Fortress: $65 million (or $100 million) loan, repaid in patents.

9. The aftermath and "Theranos 2.0"

Billy Evans, Holmes's partner, has run Haemanthus in Austin since 2022 (incorporated 2023). It says it will analyze blood, urine and saliva with Raman spectroscopy and deep learning, first for pets, then humans. A Raman patent was granted in January 2025. It has raised roughly $18 to $20 million toward a $50 million goal (investor materials described a wearable human device needing three years and $70 million), employed about a dozen people, and its chief technology officer resigned on March 16, 2026. As of November 2026 it has no product and no independently validated data. In a 2026 documentary Evans calls a medical-testing venture "Theranos 2.0" and says the idea is "to put the band back together." A source close to Evans says that project is separate from Haemanthus, and Haemanthus says Holmes has no role.

On the science: Raman spectroscopy is real and well established for chemical identification, but whole blood is one of the hardest matrices for it. Raman scattering is orders of magnitude weaker than the excitation light; blood produces strong fluorescence and hemoglobin absorption that swamp weak peaks; target molecules such as hormones are present at nanomolar to picomolar concentrations; surface-enhanced Raman (SERS) improves signal but suffers from substrate-to-substrate irreproducibility. A 2026 critical assessment in Clinical Chemistry and Laboratory Medicine lists exactly these barriers to translation. Glucose quantification in blood phantoms with liquid-core waveguides has been shown in research, which is the plausible kernel of the idea. Multiplexed disease detection from a Raman spectrum of whole blood with deep learning is unproven and, given the signal physics, unlikely to be the first thing that works.

10. Lessons, turned into rules for Thimble Dx