A Harvard-led team has built DNA on the surface of an ordinary silicon chip — using an enzyme in plain water, rather than the harsh solvents industry relies on — and made 64 different strands at the same time. It is the largest published demonstration of its kind. Set against two decades of cost data, it is a laboratory first: no price per letter has been published for it.
DNA is a chain built from four chemical units, abbreviated G, A, T and C. Each unit is called a base — think of it as one letter. All the information is in the order of the letters.3
Reading means taking DNA that already exists and working out the order of its letters. That is what a sequencing machine does. It is the half of the field everyone has heard about: the first draft of a human genome, roughly 3 billion letters, cost about $300 million and took the Human Genome Project until 2000; by late 2015 a draft of one person's genome cost under $1,500 at the centres America's National Human Genome Research Institute funds.3 That collapse is why reading is now ordinary — NHGRI attributes the sharper scrutiny of its own costs to the growing scale of human genetics studies and the growing number of clinical uses for sequencing.3 Diagnosing an inherited disease, sequencing a tumour, identifying an outbreak: all of it is reading.
Writing is the opposite operation, and it is not copying. It means manufacturing a strand from scratch whose letter order you choose, one letter at a time. Written DNA is a consumable rather than a finding: it is the ingredient inside diagnostic tests, the raw material for engineering cells and building genes, a standing tool of cancer research — and, in prospect, a way to archive digital data in molecules.17 Almost every lab that reads DNA first has to buy some written DNA to do it with.
This piece is about the writing half, and about one number: what it costs to put down a single letter.
On 17 June a Harvard-led group reported in Nature Electronics that a silicon chip had built up to 64 different DNA strands at once, each 38–39 letters long. Only 10–11 of those letters differ from strand to strand and carry the message; the rest is shared scaffolding used to anchor and retrieve them.1
The trick is control. Adding a letter requires unlocking the tip of the strand first, and acid does the unlocking. The chip's surface holds 256 tiny sites, each a pair of concentric metal rings around an anchored strand. Pushing current through the inner ring makes acid exactly there; pulling current through the outer ring mops up any acid that spreads. Do that at only the sites due for a letter, cycle after cycle, and 64 different strands grow side by side in what is otherwise ordinary water.1 The team then used the 64 strands to store a 169-byte piece of text as a demonstration — a byte being, in plain text, roughly one character.1
Two numbers make this news. Writing DNA with an enzyme in water, rather than with reactive chemicals in solvent, had previously been done in parallel only at about a dozen strands at a time.47 Sixty-four is roughly five times that. It is also about 125,000 times below what the incumbent chemistry already does on the industry's own advertised figures: one supplier markets pools of up to 8 million short strands on a single chip.11
How accurate was it? The paper's public abstract gives no headline error rate. In a companion run of 64 slightly shorter strands storing a 127-byte text, the consensus sequence at every site — the letter order you get by reading many copies of a strand and taking the most common letter at each position — matched its target exactly, and did so again when the 64 strands were shuffled to different sites.1 Errors in individual copies are mapped in the paper's extended data, but those values sit behind the journal's paywall and this desk could not read them. The same goes for how long one cycle takes, which neither the abstract nor Harvard's summary states.17
A 64-strand demonstration is news because the writing half of DNA has crawled for two decades while the reading half fell off a cliff. Between 2001 and 2017 the cost of reading one letter at NHGRI's funded centres fell from $5.29 per thousand letters to about 1.1¢ per million — a factor of 469,184.28 Over the same years the price of writing one letter, in the only long-run compilation there is, fell from $12 to 2¢ — a factor of 600.9 Reading fell 782 times further. Put another way: reading halved in price every ten months, writing every 21.
Any claim about cost rests on throughput: how many letters a machine lays down per run. Here the paper is unusually candid about its own ceiling. The team made one 10-letter sequence using all 256 sites, then made it again using only four sites spaced far apart. Sixty-four times the sites should give roughly sixty-four times the product. It gave about 4.3 times — 6.7% of what simple arithmetic predicts.1
A denser chip failed outright. The team traced the failure not to the electronics — the acid stayed where it was put — but to the chemistry that does the unlocking. Acid does not strip the cap off the strand directly; it creates go-between molecules that do, and those drift to neighbouring sites, blurring the boundary the chip worked so hard to hold.7 When neighbouring rows were switched on together, most strands came out shorter than intended and some were not detected at all.1 The authors' own conclusion is that a chemistry in which acid does the job directly would let output scale with the chip.1 That chemistry does not yet exist.
The incumbent is phosphoramidite chemistry: the standard industrial method since the 1980s, which builds DNA out of reactive chemical units suspended in organic solvent. It is very good at scale — it already writes millions of strands in parallel — but it needs hazardous solvents and centralised plant to do it.7 The published chemistry literature is plain that the method demands a substantial excess of reagents and considerable volumes of solvent, mostly acetonitrile, used less to dissolve anything than to wash the apparatus between the four steps of every cycle.5 A technical note from Glen Research, a supplier of those reagents and therefore an interested party, puts the waste at up to 3–5 mL per cycle, with roughly 7–8 mg of reagent spent in excess even at the smallest routine scale.10
That is the measured comparison, and it is the whole of it. A water-based route removes the solvent. It has not been shown to remove the cost, because no cost per letter for the chip has been published anywhere.
There is an official, documented, decades-long series for the cost of reading DNA. There is no equivalent for writing it. The write data exists because one analyst, Rob Carlson, assembled it, published it in Nature Biotechnology in 2009, and gave the field its shorthand.6 He is candid about what it is: numbers gathered by telephoning people and reading technical notes, in a market where the price of a written letter has frequently sat below what it cost to make, and where his productivity series had to stop because the surviving suppliers keep their machines proprietary.9 His cost table ends in 2017. NHGRI makes the same warning about its own half of the picture: a price may be higher or lower than the underlying cost.3
Three things, and they are the three that matter. Error rate at length. The consensus strands read back correctly at 35–36 letters1, but the per-letter error maps and the share of strands coming out full-length are in paywalled extended data this desk has not read — and nothing published establishes how errors pile up beyond 39 letters. Scale-up. The denser chip failed, and the authors' way past it — a chemistry where acid unlocks the strand directly — is a stated research direction, not a result.7 Whether the write curve can be measured at all. The only public per-letter series for writing is one analyst's survey data, self-described as limited by what he could get, priced below cost at times, and stopped in 2017.9 You cannot show that a trend has bent when the trend itself is measured this thinly.
Specialists disagree about what the demonstration implies, and the evidence to settle it does not exist yet. Both positions below are arguments about facts, not about values — which means evidence could in principle settle them. It just hasn't.
The chip did exactly what was asked of it: it held the acid where it was told to. It has 256 sites and only 64 were used, so there is headroom built in. The failure at higher density traced to the unlocking chemistry, not the electronics, and the authors' analysis concludes that a direct-acid chemistry would let output scale with the chip.17 Silicon is the one manufacturing base that has reliably scaled for fifty years.
Output already stops keeping up inside the current array: 64 times the sites returned 4.3 times the product, and firing neighbouring sites together produced stunted strands and missing ones.1 Sixty-four strands sits some 125,000 times below advertised commercial capacity11, the chemistry that would close the gap is unspecified, and no cost or speed figure has been published at all. On the record as printed, the price of writing a letter has not moved.9
Desk analysis, not fact. The public evidence settles neither case. What it does settle is narrower, and worth saying plainly: this is a demonstration that many strands can be written at once, not a demonstration that writing has got cheaper — and the chip's own output already grows more slowly than the number of sites it uses.