DEV Community: InvisibleHill

Memory Is the Bottleneck, but Not Forever

InvisibleHill — Sat, 08 Aug 2026 04:08:10 +0000

Originally published by InvisibleHill Research. This cross-post preserves the original research cut-off and source list.

Research cut-off: May 29, 2026. Market conditions and company guidance may have changed after publication.

Memory used to be the semiconductor industry’s least romantic business. It was
capital-intensive, brutally cyclical, and difficult to differentiate. Producers
spent billions of dollars making bits that customers mostly bought on price.
When supply exceeded demand by a few percentage points, profits disappeared.

Artificial intelligence has rewritten that story so thoroughly that the old
description now sounds quaint. World Semiconductor Trade Statistics expects
global memory revenue to rise about 250 percent in 2026, to more than $800
billion. Micron’s fiscal third-quarter revenue reached $41.5 billion, more than
four times the year-earlier level. It also guided to an 86 percent gross margin
for the next quarter. Sandisk’s gross margin jumped from 22.5 percent to 78.4
percent in a year.

Those are not normal semiconductor numbers. They are scarcity numbers.

The scarcity is real. AI accelerators cannot work without high-bandwidth memory;
inference servers need ordinary DRAM as well; vector databases, model
checkpoints, and retrieval systems consume flash; and the resulting oceans of
data eventually settle onto hard drives. But “storage” is not one market, and
real demand does not make every price rational. The same extraordinary profits
that prove the shortage also finance its eventual cure.

The useful investment question is not whether AI needs more memory. It does. The
question is who will still earn excess returns after customers, competitors,
and new capacity respond.

Four businesses hiding inside one theme

Investors often group Micron, Samsung Electronics, SK hynix, Sandisk, Western
Digital, and Seagate into a single AI-storage trade. Their economics are quite
different.

High-bandwidth memory (HBM) is stacked DRAM placed beside an accelerator.
Its value lies in moving enormous volumes of data without starving the GPU or
custom AI chip. HBM requires leading-edge DRAM, advanced packaging, thermal
management, a logic base die, and lengthy qualification with a small number of
accelerator customers. A failed stack can waste expensive silicon around it.
Yield and reliability matter as much as raw bit output.

Conventional server DRAM holds the working state of CPUs and accelerators.
It is less exotic, but AI servers carry far more memory than traditional
machines. It also benefits indirectly from HBM: every wafer and clean-room bay
directed toward complex HBM is capacity not used for ordinary DRAM.

NAND flash and enterprise SSDs sit farther from the processor. They store
model weights, embeddings, retrieval indexes, training checkpoints, and the
rapidly growing key-value caches created during inference. Enterprise products
earn premiums for endurance, firmware, latency consistency, and power
efficiency. Yet the underlying NAND bits remain more interchangeable and supply
is relatively elastic once factories are running.

Nearline hard drives are the archive. They are too slow for an accelerator’s
working memory but economically difficult to replace when hyperscalers need to
retain exabytes. Seagate and Western Digital now operate in what is effectively
a two-vendor market for high-capacity cloud drives. Higher areal density lowers
cost and power per terabyte, giving hard drives a durable role beneath flash.

HBM is a bottleneck, not a permanent monopoly

HBM deserves the highest-quality multiple in the group today. Its demand is
tied to the number and memory content of accelerators, not merely to replacement
PCs. Each generation is harder to manufacture. HBM4 doubles the interface to
2,048 data pins, adds a sophisticated base die, and integrates more deeply with
the customer’s system design. Suppliers cannot redirect commodity inventory
and call it qualified HBM.

The market therefore rewards execution. Micron is shipping HBM4 in volume for
its lead customer and expects HBM4E production in 2027. Samsung began commercial
HBM4 shipments in February and expects its HBM revenue to more than triple this
year. SK hynix entered the cycle with the strongest HBM franchise and continues
to sell high-value memory into AI systems.

This is a genuine moat, but it is a moving one. Product leadership must be won
again at every generation. Qualification is concentrated among a few enormous
buyers with the engineering staff and purchasing power to encourage a second or
third source. Packaging partners can add capacity. Yields improve. Samsung’s
return as a credible HBM4 supplier is good for the ecosystem and potentially
bad for scarcity rents.

Most important, HBM consumes capital. Micron expects roughly $27 billion of
capital spending in fiscal 2026, with quarterly spending rising again in 2027 as
it pulls forward clean-room construction. That expansion will not produce
finished supply immediately; fabs take years and HBM packaging remains complex.
But an 86 percent gross margin is an invitation written in very large type.

The bear case does not require AI demand to collapse. Supply only needs to grow
a little faster than customers’ urgent requirements, or accelerator makers need
to pause between architectures. In a fixed-cost industry, the marginal few
points decide the price.

Ordinary DRAM may be the most misunderstood winner

HBM attracts the headlines, but conventional DRAM explains why the current
cycle has spread across phones, PCs, automobiles, and industrial products.

In Micron’s latest quarter, DRAM bit shipments rose only in the low single
digits sequentially while average prices increased in the low-60-percent range.
That distinction is crucial. Revenue growth came overwhelmingly from price and
mix, not a sudden flood of physical bits. Samsung likewise attributed record
memory earnings partly to higher average selling prices and limited supply.

AI creates direct demand through memory-heavy servers and indirect scarcity by
pulling advanced capacity into HBM. Meanwhile, the producers that survived the
last downturn have become more cautious about adding commodity output. Micron
has signed 16 strategic customer agreements, generally running through 2030,
that provide supply assurance and may make future revenue less exposed to the
spot market.

That could represent a healthier industry structure. It does not repeal the
cycle. Long-term agreements differ in pricing, volume commitments, remedies,
and flexibility; “contracted” is not synonymous with “fixed high margin.” Node
transitions also create more bits per wafer without a new building. Once fresh
clean-room capacity arrives, ordinary DRAM lacks HBM’s packaging and
qualification barriers.

Investors should separate bit growth from price growth every quarter. If unit
economics are improving while bits remain constrained, earnings can stay
exceptional. If bits accelerate just as average selling prices flatten, the
profit peak may already be visible in the income statement.

NAND has the best demand story and the weakest memory

Enterprise flash has a compelling AI use case. Training creates checkpoints;
inference creates retrieval indexes and caches; agentic systems repeatedly read
large datasets. Micron has begun shipping a 245-terabyte QLC SSD. Sandisk’s
data-center revenue rose 233 percent sequentially in its April quarter. These
are not consumer USB drives being relabeled as AI.

But NAND has repeatedly punished investors who confuse a growing market with a
good industry. Producers can stack more layers, improve bits per cell, and turn
idle output back on. Customers can defer purchases when inventories rise.
Qualification and firmware create a moat at the drive level, but a large part of
the value remains sensitive to the price of NAND wafers.

Sandisk offers the clearest evidence. Its fiscal third-quarter revenue nearly
doubled sequentially to $5.95 billion, helped by a richer data-center mix and
higher pricing. Its 78.4 percent gross margin was 56 percentage points above the
prior-year level, and management guided to 79–81 percent for the following
quarter. The company has also signed multi-year arrangements with firm financial
commitments, which may improve durability.

The bullish reading is that AI has reset the value of enterprise flash. The
cautious reading is that no mass-produced storage medium retains an 80 percent
gross margin unless supply is extraordinarily tight. Both can be true. A
structurally larger market can still deliver violently cyclical earnings.

Hard drives: the quietest and perhaps cleanest case

The hard-drive thesis is less spectacular. It may be more legible.

After separating its flash business, Western Digital is a focused HDD company.
Its fiscal third-quarter revenue rose 45 percent year over year to $3.34 billion
and GAAP gross margin reached 50.2 percent. Seagate reported $3.11 billion of
revenue and a 46.5 percent GAAP gross margin. Both are benefiting from
hyperscale demand, disciplined supply, and higher-capacity drives.

AI strengthens “data gravity”: generated data may be processed in fast memory,
but useful records, model artifacts, video, telemetry, and backups must live
somewhere. Nearline HDD remains the cheapest practical home for much of it.
Heat-assisted magnetic recording and other areal-density gains can raise
capacity without equivalent growth in heads, media, power, or floor space.

The risk is that industry discipline is partly a product of consolidation and
long qualification cycles, not immunity from substitution. Flash cost per bit
continues to fall. Hyperscalers are a concentrated customer group. A delayed
capacity transition can create shortages; a successful one can produce more
exabytes than expected. And when a mature hardware company earns a 50 percent
gross margin, the market may capitalize a scarcity period as if it were a new
permanent baseline.

Is there a bubble?

There is no convincing evidence that AI memory demand itself is fictitious.
Shipments, customer commitments, product qualifications, and cash flow are
visible. This is not a pre-revenue theme assembled from investor slides.

There is, however, a classic bubble risk in the extrapolation.

WSTS expects the memory market to grow roughly 250 percent in 2026 and then 32
percent in 2027. Even the forecast embeds a dramatic deceleration. Micron’s
latest DRAM price increase, Sandisk’s margin expansion, and the HDD makers’
record profitability all describe an industry operating far above its
through-cycle economics. Stocks can fall while earnings rise if the rate of
improvement merely becomes less extraordinary.

Trailing price-to-earnings ratios are particularly unhelpful near a cycle peak.
The denominator is changing faster than the numerator. A producer can appear
cheap on the quarter in which scarcity is greatest and expensive one year later
without its share price moving. The better exercise is to normalize price,
utilization, and margin across a full replacement cycle, then ask what portion
of today’s demand and customer structure is genuinely new.

A disciplined investor should watch six things:

bit shipments versus average selling prices: price-led growth is powerful but more reversible;
capital spending and clean-room timing: today’s capex is tomorrow’s supply;
inventory days at producers and customers: shortages often end before management language changes;
HBM yield and customer qualification: leadership is valuable only while it remains hard to replicate;
contract quality: duration matters less than enforceable volume, pricing, prepayments, and cancellation terms; and
gross margin by product: a blended number can hide commodity exposure behind a small, premium category.

The investment conclusion

The memory boom is not one trade. HBM has the strongest near-term scarcity and
the deepest technical barriers. Conventional DRAM has the largest indirect
benefit from capacity diversion. Enterprise NAND has enormous AI volume
potential but the most dangerous history of supply elasticity. Nearline HDD
offers slower growth, a concentrated market, and a credible cost advantage for
cold and warm data.

The highest-quality companies will use this windfall to improve technology,
secure customer commitments, and return cash without assuming that today’s
prices last forever. The weakest investment cases will require both perpetual
AI demand acceleration and permanent scarcity margins.

AI has elevated memory from a component to a system constraint. That change is
structural. Constraints, however, attract engineers and capital. The winners
will not simply be the firms selling the scarcest bit in 2026. They will be the
ones whose advantage survives the moment the shortage begins to work.

Sources

This report compares operating economics rather than recommending individual
securities. Company claims are identified as such; financial figures are GAAP
unless noted. The framework is based on public information available through
May 29, 2026.

Bitcoin’s Quantum Problem Is a Governance Problem

InvisibleHill — Sat, 08 Aug 2026 04:01:02 +0000

Originally published by InvisibleHill Research. This cross-post preserves the original research cut-off and source list.

Quantum computing will not “decrypt Bitcoin.” That phrase is technically wrong
and politically distracting.

A sufficiently capable quantum computer could do something narrower and more
dangerous: derive a private key from a visible public key, then produce a valid
signature and spend coins it does not own. The blockchain would not be hacked.
Its rules would be followed with a stolen key.

No existing machine can do this. Today’s best public demonstrations are still
learning how to preserve and operate a small number of error-corrected logical
qubits. But the resource estimates are falling, post-quantum standards now
exist, and Bitcoin upgrades slowly by design. Waiting for an attack is not a
serious plan.

The cryptography is solvable. The harder problem is the migration: how to move
millions of coins, what to do about owners who never move, and whether the
network should eventually reject signatures that are mathematically valid
under today’s rules. Quantum computing is therefore not only a security threat.
It is a test of Bitcoin’s claim that no committee decides whose money is real.

Two algorithms, two very different threats

Public discussion often combines Shor’s algorithm and Grover’s algorithm into
one quantum menace. They do different things.

Shor’s algorithm attacks ownership. Bitcoin uses elliptic-curve
cryptography on the secp256k1 curve for ECDSA and Schnorr signatures. Given a
public key, a large fault-tolerant quantum computer running Shor’s algorithm
could solve the discrete-logarithm problem and recover the corresponding
private key. With that key, the attacker could sign a normal transaction.

Grover’s algorithm weakens search. Bitcoin’s proof of work and many internal
commitments depend on SHA-256 and other hash functions. In an idealized quantum
computer, Grover’s algorithm gives a quadratic speedup: a 256-bit preimage
problem has roughly 128-bit quantum security. That is a meaningful reduction,
but 2^128 remains an enormous search space. Quantum mining would also face
serial gate costs, hardware scarcity, competition from classical ASICs, and the
network’s difficulty adjustment.

The practical conclusion is clear. The first systemic quantum risk is not a
miner rewriting the chain. It is a key-recovery attacker spending coins from
outputs whose public keys are already visible.

Not every coin is equally exposed

Bitcoin has several output types, and they reveal different information.

Early pay-to-public-key (P2PK) outputs place a public key directly on-chain.
Modern Taproot outputs also contain a public key in the output. Coins in those
formats offer a future quantum attacker unlimited time to work: the public key
is visible as long as the coin remains unspent.

Pay-to-public-key-hash formats, including common SegWit addresses, work
differently. The output commits to a hash of the public key. The public key is
revealed only when the owner spends. If the address has never been reused, a
quantum attacker must recover the private key and replace the transaction in
the brief interval before confirmation.

That creates two clocks. A long-exposure attack can target an already
visible key for months or years. A short-exposure attack must observe a
transaction in the mempool, recover the key, and front-run the owner, perhaps in
minutes. The latter requires a much faster machine.

Address reuse defeats the concealment. Once a public key has appeared in a
previous spend, any remaining or newly received coins controlled by the same
key become long-exposure targets. Extended public keys and some wallet
descriptors can create exposure off-chain as well.

This nuance cuts both ways. It means the entire supply is not waiting naked for
the first quantum computer. It also means the oldest coins, reused addresses,
and Taproot outputs are natural first targets. A successful theft from a famous
dormant address could cause a market crisis before the attacker had the speed
to threaten ordinary transactions.

How far away is the machine?

Nobody can responsibly give a date for a cryptographically relevant quantum
computer. Qubit counts in headlines are poor forecasts because a noisy physical
qubit is not a reliable logical qubit, and logical qubits alone say little about
gate speed, connectivity, error correction, or how long a circuit can run.

Google’s 105-qubit Willow processor crossed an important threshold in 2024. A
larger error-correcting code suppressed errors better than a smaller one, and a
101-physical-qubit code preserved one logical memory longer than its best
component qubit.
That was a scientific milestone, not a cryptographic computer. Shor’s algorithm
against secp256k1 requires a large universal fault-tolerant machine executing a
deep sequence of reliable operations, not merely storing one encoded state.

The gap is still enormous, but it is no longer useful to describe it with a
number from five years ago. A 2026 resource study accepted by PRX Quantum
describes circuit designs using fewer than 1,500 logical qubits and estimates
that, under specific superconducting-hardware assumptions, fewer than 500,000
physical qubits could recover a secp256k1 key in
minutes.
That estimate is not a delivery forecast. It assumes error rates, connectivity,
control, and manufacturing at a scale nobody has demonstrated. It does show
why “millions of qubits” should not be treated as a permanent safety margin.

There is another reason to start early: migration time. NIST finalized its
first post-quantum standards in 2024, including the lattice-based ML-DSA and the
hash-based SLH-DSA for digital signatures. The U.S. government’s transition
planning contemplates phasing out vulnerable public-key cryptography in the
2030s. Bitcoin cannot assume that banks, browsers, and governments will spend a
decade migrating while a global decentralized asset will improvise later.

The sensible stance is neither “quantum winter forever” nor “Q-Day in 2029.” It
is to treat arrival time as a wide distribution and make migration progress
observable.

The first step exists, but it is not the finish

Bitcoin Improvement Proposal 360 is the most concrete protocol step so far. It
proposes Pay-to-Merkle-Root (P2MR), a new output type similar to Taproot but
without the elliptic-curve key-path spend. The output commits directly to a
script-tree root, so it can avoid leaving a long-lived public key exposed.

The proposal is deliberately modest. Its own specification says P2MR protects
against long-exposure attacks, not the short-exposure window after a public key
is revealed.
Full protection still requires a post-quantum signature scheme inside the
spending script.

That separation is good engineering. P2MR can reduce today’s attack surface and
create a clean place for a future signature opcode without forcing the network
to choose an immature algorithm immediately. It can be activated through a
soft fork, preserving compatibility for nodes that have not upgraded.

The difficult part is size. A Schnorr public key is 32 bytes and a signature is
64 bytes. Standardized post-quantum signatures are generally measured in
thousands or tens of thousands of bytes. Bitcoin has intentionally scarce block
space. Replacing every signature naively would reduce transaction throughput,
raise fees, enlarge the blockchain, and make node operation more demanding.

Developers are studying lattice signatures, hash-based signatures, aggregation,
commit-and-reveal constructions, and zero-knowledge proofs that verify a
post-quantum authorization more compactly. Each trades bandwidth, computation,
assumptions, implementation risk, and maturity. NIST’s selection is valuable
evidence; it is not an automatic answer for a consensus system that may need to
validate signatures for generations.

The lost-coin dilemma

Suppose Bitcoin adds a sound post-quantum output and gives holders years to
migrate. Most active users, exchanges, custodians, and institutions move. Some
coins do not.

They may belong to people who missed the deadline. The keys may be in cold
storage. The owner may be dead. The coins may be deliberately dormant. Or the
keys may be permanently lost. The blockchain cannot distinguish among these
stories.

Once a quantum attacker can recover old private keys, the network faces three
uncomfortable choices.

Honor the existing signature rules. A quantum-generated signature is
valid, so miners include it. This preserves protocol continuity but rewards
theft. Dormant coins could suddenly enter liquid supply, and no observer
could tell whether the spender was an owner or an attacker.
Disable vulnerable signatures before the attack. This protects the
network from mass theft but strands any owner who failed to migrate. A rule
intended to defend property becomes a collective decision to freeze it.
Create a recovery mechanism. Owners could prove knowledge of an old
secret using a post-quantum construction and reclaim funds after elliptic-
curve spends are disabled. This reduces confiscation risk but adds protocol
complexity, privacy questions, and new failure modes.

The argument becomes most intense around coins attributed to early miners,
including Satoshi Nakamoto. If those keys are lost, a quantum computer could
turn economically dormant supply into active supply. If the network freezes
them, critics will say developers changed the rules to protect the price. If it
does not, critics will say it knowingly legalized theft.

No cryptographic primitive can answer that question. The answer will come from
social consensus expressed through software, miners, economic nodes, wallets,
and users.

A credible migration plan

A robust plan should be staged long before an emergency.

First, measure and reduce exposure. Wallets and custodians should inventory
output types, address reuse, extended-key handling, and recovery paths. Users do
not need to panic-move coins today. They do need modern backups and software
capable of participating in a future migration. Wallet defaults should stop
creating avoidable long-exposure keys.

Second, activate an upgradeable quantum-safer output. P2MR or a comparable
design would let cautious holders move before the final signature choice is
settled. Exchanges and hardware wallets need test vectors, descriptors, fee
estimation, recovery procedures, and years of operational testing.

Third, select at least two authorization paths. A primary scheme and a
conservative backup based on different mathematical assumptions would reduce
the danger of discovering a flaw after billions of dollars have migrated.
Hybrid signatures can require both classical and post-quantum authorization
during the transition.

Fourth, publish triggers. The network needs evidence thresholds for
accelerating migration and, eventually, disabling vulnerable spends. A July
2026 developer discussion proposed a cryptographic “tripwire”: a signature by a
published nothing-up-my-sleeve key could prove that secp256k1 had been broken
and automatically disable elliptic-curve paths in new protected outputs. The
idea is not finished, but it correctly asks who gets to declare the emergency.

Finally, decide the legacy policy while the choice is still abstract. A
multi-year notice period, transparent activation sequence, and narrowly scoped
recovery design are more legitimate than an overnight patch after a theft.
The worst time to debate property rights is while a quantum attacker is moving
coins.

What this means for Bitcoin’s value

Quantum risk is unlikely to appear in the price as a smooth annual discount. It
will arrive through discrete credibility events: a new error-correction result,
a smaller resource estimate, a state program, a public key-recovery
demonstration on a weaker curve, or a successful Bitcoin testnet migration.

If protocol work stays ahead of hardware, each milestone can increase
confidence. Bitcoin would demonstrate that fixed monetary rules do not require
frozen technology. If governance stalls while the hardware gap closes, the
same milestone can trigger a rush to safer output types and a repricing of
dormant-coin risk.

An actual theft would be different. The immediate problem would not be a change
to the 21 million cap; quantum computers cannot mint arbitrary bitcoin under
current rules. It would be a sudden transfer from inactive or compromised
owners to an adversary with no acquisition cost and every incentive to sell or
destabilize the market. Confidence in custody and settlement would matter more
than the formal supply schedule.

Bitcoin has survived bugs, forks, exchange failures, and political bans because
the protocol and its users adapted without changing the core monetary bargain.
Quantum computing asks whether they can do that prospectively, before the threat
forces a choice between theft and confiscation.

That is why the work should begin while it still feels early.

Sources

This analysis separates cryptographic possibility, engineering readiness, and
network governance. Resource estimates are scenario-dependent and should not be
read as arrival-date forecasts. Technical context is current through January 26, 2026.

BIP 110 and the Cost of Policing Bitcoin's Block Space

InvisibleHill — Sat, 08 Aug 2026 03:49:09 +0000

Originally published by InvisibleHill Research. This cross-post preserves the original research cut-off and source list.

Research cut-off: July 30, 2026. Miner signaling is a live metric and may have changed after publication.

BIP 110 begins with a problem that many Bitcoin users can recognize. A miner
can collect a one-time fee for including an image, token payload, or other
arbitrary data, while thousands of node operators absorb the cost of
downloading, validating, and sometimes storing it. The proposal's authors see
that mismatch as a subsidy for data storage and a threat to Bitcoin's use as
money.

Their answer is a temporary soft fork. For about one year, BIP 110 would make
several currently valid transaction structures invalid under consensus rules.
It would cap OP_RETURN outputs at 83 bytes, limit many data pushes and witness
items to 256 bytes, restrict large output scripts and Taproot control blocks,
and disable several Taproot upgrade paths and script features that can carry
data.

Adam Back agrees with the premise more than his opposition sometimes suggests.
He has said that Bitcoin is about money and that spam has no place in its
timechain. He also designed Hashcash to make spam costly. His objection is to
the remedy. In Back's view, an annoyance that fits inside Bitcoin's existing
block limit does not justify a contentious consensus change, especially one
that can be bypassed, can interfere with legitimate scripts, and has not earned
technical or economic agreement.

That distinction is the center of the debate. BIP 110 asks whether Bitcoin
should discourage an unwanted use through local policy and fees, or declare
some forms of that use invalid for everyone. Back's case is stronger on this
question. The proposal identifies a real externality, but it offers an
asymmetric bargain: incomplete filtering in exchange for a new consensus
precedent, a more complicated upgrade path, and a material risk of a minority
chain.

A policy dispute became a consensus proposal

The immediate background is Bitcoin Core 30.0. Released in October 2025, it
raised the default -datacarriersize setting to 100,000 bytes, effectively
removing the previous OP_RETURN relay limit because the standard transaction
size limit is reached first. It also allowed multiple data-carrier outputs
within one transaction. Operators can still set the limit back to 83 bytes.

This was a change in node policy, not Bitcoin consensus. Policy determines
which valid unconfirmed transactions a node will relay and which valid
transactions a miner may place in its candidate block. Different nodes and
miners may choose different policies. Consensus determines whether a confirmed
block is Bitcoin at all. A transaction filtered by one node can still reach a
miner through another route, and a block containing it remains valid to the
network.

BIP 110 crosses that boundary. Formally titled the Reduced Data Temporary
Softfork, it would add seven consensus restrictions for one year:

New output scripts could not exceed 34 bytes, except for an OP_RETURN output of up to 83 bytes.
Many pushed data fields and script-argument witness items would be limited to 256 bytes.
Spending undefined witness or Tapleaf versions would be invalid.
Taproot annexes would be invalid.
Taproot control blocks would be limited to 257 bytes.
Tapscripts containing an OP_SUCCESS opcode would be invalid.
Tapscripts executing OP_IF or OP_NOTIF would be invalid.

Outputs created before activation are grandfathered. After the temporary
period expires, the extra rules disappear. The proposal reached "Complete"
status in the BIP repository on June 25, 2026, which means its specification is
complete. It does not mean Bitcoin has accepted or activated it.

The deployment is unusually assertive. Early lock-in requires 1,109 of 2,016
blocks, or 55 percent, to signal with version bit 4. If that threshold is not
reached voluntarily, BIP 110 nodes enter a mandatory signaling period at block
961,632. During that period they reject non-signaling blocks. Lock-in follows
no later than block 963,648, with the new transaction rules scheduled to become
active at block 965,664.

At 15:35 UTC on July 30, the public BIP 110 monitor showed 14 signaling blocks
out of 655 in difficulty period 476, a rate of 2.14 percent against the
55 percent target. That snapshot does not determine the final outcome, but it
shows how far the proposal remained from coordinated miner activation shortly
before mandatory signaling.

The strongest case for BIP 110

Reducing the proposal to a fight over JPEGs misses its best argument.

Bitcoin block space has a hard supply. Data transactions compete with payments
for that space and can raise fees when demand is high. The miner receives the
fee, while every validating node must process the block. Archival nodes retain
the data, and pruned nodes must still download and validate it during initial
sync. Large spendable output scripts are more costly because they can remain
in the UTXO set, which nodes need to query quickly.

The cost is bounded by Bitcoin's block-weight limit, and pruning reduces
storage requirements. It is not imaginary. One party can buy permanent
publication while distributing part of the bill to infrastructure it does not
pay. BIP 110 supporters are also right that a fee market does not compensate
every node operator. It compensates miners for block production.

The proposal has several thoughtful safeguards. It is temporary rather than
permanent. It preserves pre-activation UTXOs. It provides test vectors and a
reference implementation. Its limits are simple enough to audit. It openly
admits that BitVM, Miniscript constructions, future upgrade hooks, and some
experimental Taproot uses may be constrained. That is better than presenting a
filter as costless.

BIP 110 also does not claim that steganography can be eliminated. A determined
user can split data into smaller pushes or disguise it inside structures with
another protocol meaning. The authors' narrower goal is to make contiguous,
obvious publication more cumbersome and to state that arbitrary data storage
is unsupported.

That may work as a social and economic deterrent. Raising the engineering cost
of an activity can reduce casual use even when it cannot stop a determined
actor. A temporary rule could also buy time for a cleaner design that separates
optional data from the material every node must validate.

The difficulty is that this modest technical effect is paired with a much less
modest activation mechanism.

Adam Back's boundary: dislike is not a consensus rule

Back's position is useful because it does not depend on liking inscriptions,
Runes, or other data-heavy applications. He dislikes spam. In September 2025,
he wrote that Bitcoin is owned by humanity, that protocol developers are
stewards, and that material changes require user consensus. By February 2026,
he was arguing that BIP 110 would damage Bitcoin's credibility as a store of
value and secure monetary network. In the same exchange he called spam "just
an annoyance."

His longer July argument starts from permissionless money. The decentralization
needed to create bearer money also prevents one faction from imposing its
preferences on everyone else. A user can change personal software, but not
another user's software. Back therefore describes BIP 110 as an attempt to
police other people through consensus, even if its supporters believe they are
defending Bitcoin.

He also emphasizes the engineering process. Bitcoin's change resistance
depends on much more than a high miner threshold. It includes review by developers, protocol
researchers, wallet authors, miners, exchanges, custodians, and users who can
refuse a change. Back compares this to an IETF-style technical consensus in
which unresolved objections block progress. In July he summarized the outcome
plainly: "Bitcoin respectfully says no to what you want."

That sentence is rhetoric, not a formal vote. Back does not own Bitcoin, and
neither do Bitcoin Core developers. His argument should stand on its mechanics,
not his biography. Here the mechanics support him. BIP 110 has weak miner
signaling, no demonstrated coalition of major exchanges or custodians, and
unresolved technical objections. The proposal's mandatory path does not create
consensus. It schedules a test of whether consensus exists.

Back's position also explains why opposing BIP 110 is compatible with running
stricter relay policy. A node operator can filter locally. A mining pool can
decline transactions. Bitcoin Knots can offer different defaults from Bitcoin
Core. Those choices compete without declaring the other side's blocks invalid.
This diversity is less satisfying to anyone seeking a universal ban, but it
preserves a valuable boundary between individual judgment and shared validity.

The technical exchange is asymmetric

BIP 110's filtering is intentionally incomplete. In October 2025, Peter Todd
demonstrated a transaction that embedded the text of the proposal while
remaining compliant with its rules. Later examples showed that larger images
could also be encoded through permitted structures. These demonstrations do
not prove that the proposal has zero effect. They do prove that arbitrary data
will survive activation.

The users most motivated to evade the filter can update their encoding
software. Ordinary wallet and contract users may have a harder problem. BIP
110 disables tools that were reserved for future upgrades and restricts
features that are not inherently data storage. It forbids Taproot annexes,
OP_SUCCESS opcodes, undefined witness versions, deep control blocks, and
executed conditionals in Tapscript. Some are unused today. Others are building
materials for future protocols.

The proposal itself acknowledges the tradeoffs. A 257-byte control-block limit
can complicate BitVM and other large script trees. Current Miniscript compilers
can produce Tapleaves containing OP_IF. No new soft fork can use the disabled
upgrade hooks during the one-year window. Wallets and libraries must understand
one rule set during the intervention and another after it expires.

Grandfathering protects outputs confirmed before activation, but it cannot
cover every pre-existing commitment. A user may hold a pre-signed chain of
transactions in which a descendant confirms after activation. To BIP 110
nodes, that descendant creates a new UTXO even though the owner committed to it
earlier. If its only spending path violates the temporary rules, the funds can
become unspendable until expiry. In the Bitcoin development mailing list,
Gregory Maxwell pointed to inheritance plans and other pre-signed arrangements
as plausible examples.

The BIP calls such cases unlikely and imposes several conditions before funds
can be frozen or lost. That may be accurate. It is still the wrong comparison
to ask whether the number is probably small. The comparison is between the
expected benefit of a bypassable filter and the expected cost of changing the
validity of unknown, privately constructed spending arrangements.

Back's critique is compelling because the asymmetry points in the wrong
direction. Sophisticated data publishers route around the restriction.
Legitimate users cannot route around a consensus rule after their spending
conditions have been fixed.

The activation design carries the largest immediate risk

A soft fork is backward compatible only when the economic majority follows the
more restrictive chain. Old nodes will accept blocks produced under the new
rules. Upgraded nodes will reject blocks that violate them. If most miners
enforce the restriction, both groups stay on one chain.

If only a small minority enforces it, the result is different. At the mandatory
signaling height, BIP 110 nodes reject blocks that the rest of the network
accepts. Unless enough hash power follows them, those nodes move onto a
low-hashrate minority chain. Transactions may be valid on both chains, deposits
become difficult for exchanges to credit safely, and users face replay and
double-spend risk. The one-year expiration does not make that coordination
problem temporary in any reassuring sense.

Supporters compare the plan with the 2017 user-activated soft fork campaign for
SegWit. The analogy is incomplete. SegWit had years of engineering review,
broad wallet and infrastructure integration, and overwhelming economic support
before the threatened UASF date. Miners ultimately activated it through the
existing signaling process, so the UASF did not have to split the network.

BIP 110's 55 percent threshold is much lower than the 95 percent used by
traditional BIP 9 deployments, and the mandatory period is meant to force the
issue if voluntary signaling fails. A threshold can coordinate actors who
already want the same change. It cannot manufacture demand from miners,
exchanges, wallets, and holders who have not agreed to bear the fork risk.

This is where node counts can mislead. Reachable nodes are not ballots, and
creating many nodes is cheap. What matters during a fork is the economic weight
that miners need to serve: exchanges that value coinbase rewards, merchants
that accept payments, custodians that control deposits and withdrawals, and
holders willing to price the two assets differently. BIP 110 has generated a
visible movement. It has not yet demonstrated that coalition.

What BIP 110 changes even if it fails

BIP 110 is unlikely to make arbitrary data disappear. Its more durable effect
will be a clearer map of Bitcoin's governance limits.

First, policy diversity will matter more. Bitcoin Core 30's OP_RETURN default
helped turn an implementation choice into an identity dispute. Bitcoin Knots,
mining-pool templates, and operator settings provide a way to express stricter
preferences without a consensus fork. Future client teams may treat policy
defaults as politically consequential even when they do not alter consensus.

Second, a failed activation would raise the bar for the next UASF. The lesson
would not be that users are powerless. It would be that user power becomes
effective only when it is connected to wallets, businesses, liquidity, and
hash rate. A social-media node count cannot substitute for economic
coordination.

Third, the technical questions will remain. Node operators still bear costs
that block-space buyers do not fully pay. Pruning, faster initial sync,
assumeUTXO, more efficient validation, optional data transport, and better
policy tools can reduce that burden without classifying transactions by
purpose at consensus. None is a perfect answer. They attack the cost directly
instead of treating cultural disapproval as a script rule.

The fee market will remain contested as the block subsidy declines. Data
applications can crowd out low-fee payments during bursts of demand. They can
also provide transaction fees to miners when monetary settlement demand is
weak. BIP 110 chooses monetary purity over that source of security revenue
without showing that payments alone will reliably replace it. The ecosystem
needs evidence about sustained fee demand, node costs, and user behavior more
than it needs a permanent label for spam.

Finally, the precedent matters. If Bitcoin changes consensus rules to suppress
one legal but disliked use, future campaigns can argue that other controversial
transactions deserve the same treatment. Governments and private pressure
groups would learn that transaction eligibility is negotiable. Back's
neutrality argument is strongest here. Censorship resistance is credible
partly because participants can say that they lack the practical consensus to
make selective exceptions.

Why policy is the better layer

BIP 110 deserves a more serious answer than mockery. It documents a genuine
cost externality, gives node operators a political voice, and forces the
ecosystem to confront what scarce block space is for. Its temporary design and
UTXO grandfathering show that the authors tried to contain collateral damage.

The proposal still fails its own risk test. It cannot eliminate arbitrary data.
It interferes with legitimate and future script uses. It creates edge cases for
pre-signed transactions. Its changing rule set burdens wallet and protocol
developers. Most of all, its mandatory activation path can split the network
without evidence that the economic majority wants the change.

Adam Back's opposition is persuasive because it keeps the layers separate. He
can dislike spam, support money as Bitcoin's primary purpose, and still reject
a consensus filter. Local policy lets an operator refuse a transaction without
requiring every other operator to share that judgment. Consensus decides which
blocks belong to Bitcoin's ledger.

Bitcoin may eventually adopt narrower limits or a better architecture for
optional data. Such a proposal would need to reduce measurable node costs,
preserve upgrade hooks, protect existing spending arrangements, resist trivial
bypass, and earn broad technical and economic support before activation.
BIP 110 does not meet that standard.

Its longer legacy may be useful anyway. Bitcoin's participants plainly do not
agree on acceptable behavior. The network protects neutrality by making any
faction's preferred restrictions unusually difficult to impose on everyone
else, including when that faction believes it is protecting Bitcoin.

Sources

This analysis distinguishes relay and mining policy from consensus validity.
It treats the BIP specification, code-linked test vectors, Bitcoin development
mailing list, Bitcoin Core release notes, and Adam Back's own public statements
as primary sources. Signaling data is a point-in-time snapshot, not a forecast.
Technical claims are current through July 30, 2026.