Semiconductor Patents Are Stuck Behind a Backlog That Built Up Years Ago

Our first pendency post found something specific: across all US utility patents, median time from filing to grant fell to a pandemic-era low in 2021, then climbed back to a decade high by 2025 — and semiconductor patents (CPC class H01L) got hit far harder than everything else, nearly doubling in pendency since their 2018 low. Worth saying up front: the broad up-then-down-then-up shape we found lines up with the USPTO's own pendency dashboard (uspto.gov/dashboard/patents/pendency.html), which shows the same pattern in its official first-action and total pendency series. That part isn't a surprising or contrarian claim — it's consistent with what the agency publishes itself.

What the dashboard doesn't show is a technology-area breakdown, and that's where the real question was: why did semiconductors specifically diverge so sharply from the industry-wide trend? Our first post ruled out the two most obvious explanations — prosecution didn't get harder (rejection and RCE rates for H01L actually fell steadily through the period), and it wasn't a filing surge (H01L's share of newly published patents fell, not rose). We closed that post with the driver still an open question. This is the follow-up.

We also ran into the same claim you might have seen

Around the same time we were digging into this, USPTO's own social channels were promoting a queue-reduction narrative — new examiner hiring, a dedicated "Pendency Oversight Strategy Team" launched in 2024, and a projected drop in unexamined inventory from about 800,000 applications to roughly 650,000 by FY2028. That's a real initiative, and FY2023-to-FY2024 total pendency did tick down slightly (20.5 to 19.9 months) in the agency's own reporting. But it's a forward-looking target, not a completed turnaround, and it's corpus-wide — it says nothing about whether any specific technology area is moving with the average or against it. Our data says semiconductors are moving against it, and by a wide margin.

The delay is entirely in the queue, not the back-and-forth

We split total pendency into two pieces: time from filing to the first office action (how long an application sits before an examiner even looks at it), and time from first office action to final grant (the actual back-and-forth of prosecution). To track this without distortion from unresolved cases, we measured it by the year the first action was actually issued — for the applications that got a first office action in a given year, how many months had they already been waiting since filing. For H01L, that wait held under 14 months from 2016 through 2020, then rose sharply: 16.1 months in 2021, 20.3 in 2023, and 29.6 months by 2025. The second stage — office action to grant — barely moved, staying in the 10-to-12-month range the entire time. Nearly all of the added delay is happening before an examiner ever picks up the case.

That's the opposite of what "harder cases" or "more amendment rounds" would look like. It's also a sharp reversal from where semiconductors started: through 2020, H01L applications were getting first office actions noticeably faster than the all-patent average (13.2 months vs. 17.7 in 2020). By 2024, semiconductors had crossed over and were running slower than the industry-wide average, and the gap kept widening into 2025 (29.6 months vs. 26.8).

One specific version of "harder cases" is worth ruling out explicitly rather than waving at: maybe existing semiconductor cases aren't sitting in prosecution longer, but each one is consuming more of an examiner's time per round — extra hours spent per office action, even if the round count and elapsed time look normal — which would leave less examiner capacity for picking up new applications. We can't measure hours directly, but we can measure two things that would move if this were happening: how many office actions a case actually goes through, and how often it ends in an appeal (a sign the examiner and applicant are struggling to agree). Neither moved the way the hypothesis predicts. Average office actions per resolved semiconductor case fell from 1.82 (cases picked up in 2016) to 1.39 (cases picked up in 2024) — tracking the all-patent decline (1.97 to 1.46) almost exactly, slightly below it throughout. Appeal rate for semiconductor cases fell even faster than the all-patent rate over the same window (1.1% down to 0.1%, versus 2.6% down to 0.3% corpus-wide). And first-action-to-final-disposition time — this time counting abandonment as well as grant, not grant alone, so a case dragging on before being abandoned would show up here — stayed flat at 9-11 months for both semiconductors and the corpus, moving together the whole time. However we cut it, semiconductor cases have been getting faster to resolve once an examiner starts them, not slower or more effort-intensive.

A wrinkle we had to control for first: CPC split H01L into new subclasses

Before testing why the queue backed up, we hit something that could quietly wreck any year-over-year comparison of H01L specifically: starting around 2022-2023, the patent classification system split several types of semiconductor devices out of H01L into new CPC subclasses — H10K (organic semiconductor devices), H10D (transistors and diodes), H10B (memory devices), H10F (photovoltaics), and a few smaller ones. That's not a change in what's being invented; it's a change in what label it gets. Only about 22% of documents carrying an H10-series label also still carry a legacy H01L label — meaning a real, growing slice of "semiconductor devices" has been moving to a new code that an H01L-only query simply stops seeing. We reran the queue-delay numbers above using the broader definition — H01L or any H10 subclass, counted once whether a document carries one label, the other, or both — to make sure the reclassification wasn't inflating the trend. It wasn't: the combined semiconductor category shows the same pattern, essentially unchanged (queue delay 11.0 months in 2016 rising to 29.3 in 2025, versus 11.0 to 29.6 for H01L alone). We used this broader, reclassification-robust definition for everything that follows.

Ruling out the two obvious "examiners" theories

A growing queue with flat prosecution difficulty points to examining capacity, so that's where we looked first — twice. First theory: fewer examiners working semiconductor cases. We counted distinct examiners issuing at least one semiconductor first office action each year: 2,438 in 2016, peaking near 2,600 in 2019, dipping to about 2,060 in 2021, recovering to 2,329 by 2025 — a dip and recovery, not a sustained decline. Second theory: the same examiners spending more time per existing case, leaving less capacity for new ones — which would show up as more office actions per case or more appeals, even without slower elapsed time. Neither showed up: office actions per resolved case fell for semiconductors (1.82 to 1.39, cases entering prosecution 2016 vs. 2024) at least as fast as the all-patent average (1.97 to 1.46), appeal rate fell faster (1.1% to 0.1% vs. 2.6% to 0.3% corpus-wide), and first-action-to-final-disposition time — counting abandonment as well as grant, so a case dragging on before being abandoned would still show up — stayed flat and matched the corpus. Examiners aren't disappearing, and they aren't getting bogged down. So if the workers and the work are both normal, what's actually causing new applications to wait longer?

The real mechanism: filings outpaced examining capacity for several years, and the backlog never fully drained

The answer is simpler than either theory above, and it's arithmetic rather than a change in anyone's behavior. We compared how many semiconductor applications were filed each year against how many actually got a first office action that same year. From 2017 through 2019, and again in 2021-2022, filings ran 20-35% ahead of first actions issued — meaning tens of thousands more applications entered the queue each year than examiners could clear. Those unresolved applications don't vanish; they accumulate as a backlog stock, and every year that inflow outpaces outflow, that stock gets larger.

That gap compounds into a growing pile of unresolved applications — from roughly 6,000 in 2014 to about 61,000 by 2022. A newly filed application doesn't just wait for an examiner to become available; it waits behind everyone who filed earlier and is still stuck in that pile. That's why the wait time for a case filed today keeps climbing even in years when filings and examiner capacity are roughly balanced (as they were by 2020-2021) — the delay isn't about this year's flow, it's about the size of the queue that built up in the years before. It also explains something that looked odd on its own: filings-per-examiner specifically for semiconductors rose from about 11 to nearly 13 over 2014-2022, while the same ratio corpus-wide barely moved (43 to 42) — a real, if moderate, capacity mismatch specific to semiconductors, small enough not to show up as an obvious productivity problem, but large enough to compound into years of backlog once it ran for several years running.

What we still don't know

We now know the queue delay is a backlog-stock problem, not a slowdown in how fast examiners work or how hard cases are to resolve. What we can't yet say is why semiconductor filings specifically outpaced examining capacity starting around 2017 — whether it's a filing surge in raw application count, an increase in something not visible to filing counts alone (e.g. this window overlaps rising semiconductor R&D investment, though we haven't tied that to a specific causal channel), or hiring/capacity planning that didn't anticipate it. We also don't know how much longer it will take for the backlog to fully drain, since that depends on whether examining capacity is now growing faster than new semiconductor filings — which our data can watch but not predict.

Why this matters if you're filing

If you're filing a semiconductor patent, the multi-year wait you're likely experiencing isn't about your application being unusually complicated, facing more scrutiny, or landing on an overloaded examiner — the data rules all of those out. It's that you're in line behind a backlog that started building years before you filed. That backlog is a real, measurable stock, and draining it takes years of capacity outpacing new filings, not a single hiring push. Budget accordingly for licensing timelines, funding milestones, or foreign-filing decisions that depend on an issued US patent, and watch USPTO semiconductor-specific first-action data (not just headline pendency) for the point where outflow durably exceeds inflow — that's the leading indicator that the wait will start coming down.