The price of frontier AI capability is falling towards zero. The consequence for an IP strategy is not that AI has become cheaper. It is that anything you thought you were buying at the model layer, you are not buying exclusively.
In July a Chinese lab released a model, free to use, that appeared to rival the paid frontier models from the large US developers. When the weights were published a week later, they came under the developer’s own bespoke licence rather than a standard open-source one. American companies had one less reason to pay for a model, and Washington had a problem it has not resolved. Part of the response has been a pre-release security review for frontier models, which one former official described as a licensing regime for frontier AI in all but name, on top of earlier measures aimed at stopping Chinese developers training on the outputs of US models.
Strip out the politics and a control-point question is left, and it is the same one that arrives whenever a capability becomes broadly available. If everyone can run the model, what is left to own?
Four IP strategy questions for anyone buying AI
– Are you paying for exclusivity at the model layer? You are not getting it. Price and term your AI contracts on the assumption the underlying capability keeps getting cheaper. Short terms, benchmarked pricing, and a right to take your prompts, evaluations and fine-tuned weights with you.
– Have you read the licence, rather than the label? Open weights are not open source, as I set out in When Everyone Can Run the Model. Bespoke licences carry attribution triggers, revenue thresholds, field-of-use limits and acceptable-use terms, and all of them travel with the weights into your product. That is a freedom-to-operate question for your business, not a procurement footnote.
– What happens if the model becomes unavailable? Export controls, release reviews and pressure not to use particular vendors make the right to deploy a regulatory variable rather than a commercial one. Any architecture only one model can serve is a dependency, and it belongs in the risk register with a substitution plan beside it.
– What have you captured at the layers that do not commoditise? Your proprietary data, the workflow your customer actually works in, your evaluations and your domain-specific tuning. In most companies these exist as undocumented practice rather than as filings, controlled trade secrets or contract terms.
None of that is new, and none of it is specific to AI. Value moves to whatever stays scarce when the headline capability does not, which is the argument in AI Isn’t Your Advantage—Your IP Strategy Is. So the question for your next strategy review is not which model is best this quarter. It is what your business still owns on the morning the model you rely on becomes free to everyone, or unavailable to you.
Read the article: China’s AI models have Trump’s AI world at war with itself, James O’Donnell, MIT Technology Review.
The patent is rarely where the money is decided. The licence is. Three IP developments from the past week make that point, and each turns on terms settled, or left unsettled, long before anyone reached a court.
In the United Kingdom, the government has published its response to its consultation on standard essential patents, the patents behind 5G, Wi-Fi and similar standards. It has dropped the proposed rate-setting track and an official essentiality-checking service, and it will not restrict injunctions. What survives is work on an early, court-anchored FRAND determination, a possible IPO opinion service and good-faith negotiation guidance. For any business that builds on standards, the signal is plain. No regulator will set your royalty. The rate will be reached by negotiation, backed by a court, and the party with the better-documented conduct starts ahead. The UK courts were already the venue to watch, as Tesla’s FRAND appeal showed.
In Delaware, a jury found that Arm breached its architecture licence with Qualcomm and interfered with Qualcomm’s relationship with Meta, then awarded two dollars. The headline is not the number. The breach finding opens the question whether Qualcomm can suspend royalties for five years under a clause Arm says is unenforceable. The judge decides that next. One obligation about software patches and tools may move billions. A licence is a set of promises, each with a price attached, waiting for the day one is tested.
In pharmaceuticals, Liquidia agreed to pay United Therapeutics a net US$380.7 million after losing on one patent covering inhaled treprostinil for a single lung condition. It will not appeal, it is removing that indication from its label, and it has given covenants that run to 2042. Its shares rose about 40 per cent. The market paid for certainty, bought with structure: one indication surrendered, the larger PAH market kept, payments staged against sales. For the patent owner, one well-aimed claim on a specific use did the work of a portfolio.
What this means for your IP licensing strategy
Value sits in the terms, not the grant. Three checks follow.
– Read your licences for the clauses that would be litigated: what counts as breach, what happens to royalties when it occurs, and which fields are in and out.
– If you implement standards, document your negotiation conduct now. It is the record a court will read.
– Settle with structure. The question is not what you pay but what you keep, for how long and in which markets. Where Is Your Leverage? makes the same point from the enforcement side, and The Reverse Payment You Didn’t Know You Made shows what a badly structured settlement costs later.
An IP strategy that stops at the grant has done half the job. The other half is the deal that turns the right into revenue, or into a bill.
Read the sources: Government sets out next steps on standard essential patents, UK Intellectual Property Office; Delaware jury awards Qualcomm $2 in damages, finds Arm breached license agreement, CNBC; Liquidia Announces Global Settlement Resolving All Pending Litigation with United Therapeutics, Liquidia Corporation.
If AI collapses the cost of making your product, the question is no longer how cheaply you can make it. It is what stops a competitor making the same thing. That is a trade secret and patent question, and most AI strategies never get to it.
A Bain & Company analysis puts a number on what is at stake. It models US$4.7 trillion of profit moving across the economy between 2025 and 2035. Only about a quarter of that is productivity. The rest is innovation and market share changing hands. The authors’ framing is blunt: “The Internet was a distribution technology. AI is a production technology.” The internet changed how products reached customers. AI changes how the product itself is made – the analysis, the diagnosis, the code, the physical work.
Their prescription is to move early, build data no one else has, and rewire workflows so that “the data you build, the signals you watch, and the recursive improvement that compounds with every interaction become proprietary intelligence.”
Proprietary only if you keep it that way
“Proprietary” is doing a lot of work in that sentence. Data and workflows are not owned in the way a building is owned. They are protected, if at all, as trade secrets – and a trade secret only exists while it is actually secret and you can show the steps you took to keep it so. The moment the profit shift starts, three things test that:
– Vendors and platforms. Your AI workflow runs on someone else’s model, cloud and tooling. Whether the data and the improvements it generates stay yours is decided by the contract terms, not by the fact that you generated them.
– People. The engineers who rewired the workflow can leave. What they carry in their heads is not a trade secret unless the documented process, and the confidentiality obligations around it, say otherwise.
– Copying. Bain’s “Rewired” cluster – pharma, healthcare, aerospace, enterprise software – is where incumbents and new entrants fight over US$1.5 trillion. In those sectors, a genuinely new AI-enabled method of producing a result is often patentable. A patent is the only right that stops a rival who reaches the same method independently.
The strategic move is to decide, for each piece of the new production layer, which protection fits: secrecy for the data and the operating know-how that cannot be reverse-engineered, patents for the methods that can. I set out that choice in Patent or trade secret? Match your IP protection to how long the value lasts. The productivity quarter of Bain’s number is the part every competitor gets too, which is the trap described in Is Your AI Strategy Just a Faster Way to Stand Still?. The other three quarters go to whoever owns something the others cannot copy – the argument in AI Isn’t Your Advantage – Your IP Strategy Is.
For a board, the question to add to the AI paper is simple. Of the data, workflows and methods this investment will create, which will we be able to stop others using, and what have we done to make that true.
Read the article: AI Will Shift $4.7 Trillion in Profits. What’s Your Stake?, Bain & Company.
There are two ways to hold a position in a technology. You can own the technology. Or you can own what it takes to build and install the technology at volume. They are not the same asset, and only one of them is scarce.
China spent August publishing the second kind. A five-year plan for oil and gas targeting a broad expansion of domestic production, pipeline infrastructure and storage. A five-year plan for coal that treats the fuel as a foundational energy source and an energy-security guarantee, pledging a peak sometime in 2026 to 2030 without naming the year. Gas elevated to an indispensable fuel for hard-to-electrify industry and, in one industry outlet’s phrase, a stabiliser for the power system. Alongside all of it, the continued build-out of the lower-carbon alternatives, and a leadership speech acknowledging that extreme weather is now a risk to the infrastructure itself.
Read as climate policy, that looks contradictory. Read as industrial strategy, it is coherent and it is aimed at a control point. Secure the inputs, own the pipelines, the grid and the factories, and you decide the rate at which anything gets deployed – whoever invented it.
What this means for your patent filing strategy
Three disciplines follow, and none of them are specific to energy.
– File where the thing is made and installed, not only where it is sold. A patent works only where it is in force. Manufacturing capacity relocates faster than most portfolios are reviewed, and a technology built and deployed in a country you never filed in is a technology you have donated. This is the jurisdictional half of the argument I made in Capacity Is Not a Moat.
– Protect the process, not just the product. Yield, throughput, tooling, materials handling and installation method are what actually limit scale, and in most companies they exist as undocumented practice rather than as filings or controlled trade secrets. That is a decision by default.
– Read your supply, joint venture and offtake contracts as IP instruments. Process improvements, tooling, field data and service rights are usually allocated in schedules nobody in the IP function has read. Give them away and you keep the patent while your counterparty keeps the market.
Timing decides how much of this is still available. Rights around an infrastructure build-out are cheap before the capacity is committed and unbuyable afterwards, which is the cost of waiting I set out in Why “Wait and See” Is Becoming the Most Expensive IP Decision You Can Make.
The prize is the position Google holds in AI compute, where rivals must keep coming back for the underlying technology, as I described in The Cash Behind the Compute. So the question for your next board paper is not whether your invention is better than the alternatives. It is who will be able to build it at scale, in which country, under whose contracts – and what you own in that chain.
Read the article: China Briefing 20 August 2026: Oil and gas plan | Xi on climate change | Coal five-year plan.
Adding AI to a product is not a strategy. Moving to where the value is going is. And every real move changes what your company controls, which makes it an IP strategy decision before it is a technology one.
An article in I by IMD puts this well. It maps AI strategy on two axes: how much of the customer relationship you control, and how many technologies and partners you have to orchestrate. Four positions follow – focused differentiation, vertical integration, collaborative ecosystem and platform leadership. The authors’ line is that “only genuine transitions – changing what a company integrates or what it controls – reposition it where value is heading.” An AI-enabled kettle is a feature. A hearing aid that becomes a health-monitoring platform is a move.
Their examples show where the IP sits. Logitech is partnering with the major AI platforms rather than building its own stack, and positioning itself as the interface between people and AI. Signify moved from selling bulbs to selling software and data relationships with enterprise customers. TomTom is the cautionary case: navigation moved into phones and cars, and the position went with it.
What the partnership licence decides
Each of those moves turns on rights that are usually settled in a partnership agreement nobody outside legal reads closely. Three questions decide whether a company controls its new position or merely supplies it:
– Who owns the interface layer? If your product is the way customers reach a platform, the protectable value is in that interaction – the design, the software, the patents on how it works – not in the commodity hardware behind it.
– Who owns the data and the customer relationship? Signify’s shift only pays if the data terms and the enterprise contracts sit with Signify, not with a platform partner.
– What happens when the platform partner enters your market? The article notes that platform owners can move into hardware directly. Your partnership terms, and the patents and trade secrets you have kept out of the joint work, are what you have left on that day.
Standards are the sharpest version of this. The article compares platform leadership to USB becoming the standard for connecting devices. Helping to set a standard is valuable only if you hold rights in it, or at least secure licences on terms that survive the standard’s success. Otherwise you have built a road for everyone else to drive on.
The same logic runs through Own the Platform, or Pay to Use It and Who Gets Paid for the Platform? Lessons From a Big Week in AI and Intellectual Property: the layer underneath the product is where the returns accrue. And the integrate-or-partner choice itself is a licensing decision, as Build, Buy or Licence? What Bayer’s Hybrid Wheat Deal Teaches About IP Licensing Strategy shows.
For a board, the useful question is not “what is our AI strategy”. It is “after this move, what will we control that we did not control before, and what do the contracts say about keeping it”.
Read the article: AI strategy is a journey – these are the levers that matter, Cyril Bouquet, Nicolas Chauvin and Julian Nolan, I by IMD.
A patent protects a mechanism. It does not protect a share of the benefit. Those are different things, and the gap between them is where a great deal of intellectual property value quietly disappears.
Cardiology supplied the evidence this month. AstraZeneca and Ionis ran a 1,432-patient Phase 3 trial of a transthyretin silencer in ATTR-CM, a progressive heart muscle disease. The mechanism works. The trial failed. Around 57 per cent of patients were already taking an older oral stabiliser when they enrolled, and roughly a quarter more started one during the study. In those patients the silencer added nothing measurable. A pre-specified subgroup taking the silencer on its own did show a benefit, but the primary endpoint was gone, and a competitor left a contested billion-dollar market.
Nothing was wrong with the patents. The incumbent standard of care had already captured most of the benefit the new mechanism had to offer. A differentiated technology is an option, not a position, which is the same lesson I drew from India’s electrification build-out in Capacity Is Not a Moat.
What this means for your patent strategy
Three disciplines follow, and none of them are specific to pharmaceuticals.
– Value your rights against the market’s default, not against nothing. Your exclusivity is worth the increment you add to what the customer is already using. Where the incumbent captures that increment, the patent is still valid and the position is still empty.
– Treat combination and sequencing rights as assets in their own right. Claims to the combination, the regimen and the order of use, and the contractual right to generate data alongside the incumbent’s product, are secured years before the evidence exists. They are cheap then and unavailable later.
– Read your evidence plan as an IP decision. The comparator you choose decides which claims your data can support and which part of the market your patents can actually reach. In pharmaceuticals that is trial design. Elsewhere it is the benchmark, the reference architecture or the interoperability target.
There is a mirror image worth holding in view. Sanofi lost the mRNA product race and can still tax the winners, because it owns the delivery platform they all need, a posture I examined in Who Gets Paid for the Platform. One company owned a mechanism the market could route around. The other owns a position the market cannot.
Assets do not lose value on their own. The board they are played on moves, which was the argument in Your patents didn’t lose value. The board where you play them changed. So the question for your next portfolio review is not whether your patents cover your technology. It is whether they cover the position your technology will be judged from, against the incumbent your customers have already chosen.
Read the article: In autopsy of failed heart disease study, AstraZeneca raises broader questions about silencer drugs, Andrew Joseph, STAT, 28 August 2026.
A technical lead is worth what you can prove. Being first, or fastest, or best matters only if the evidence behind the claim survives a hostile reading, whether from a court, an examiner, an acquirer’s due diligence team or a competitor’s expert. That is where a lot of patent strategy quietly fails, and it is the lesson in an unlikely place: a quantum computing result.
The result was reported at the end of August. IBM and the University of Chicago ran 70 error-corrected logical qubits through a sampling problem in about 15 minutes, a task the team says would take leading classical methods an impractical length of time. The distinction from earlier claims is in the second half of the announcement. The circuit was designed so that errors could be detected during the run, giving statistical evidence that the answer was trustworthy, not just fast. Quantum results have been announced before. Verifiable ones are new.
Even so, the claim was challenged within days. Other researchers questioned how far the result could be pushed and argued that advantage is an ongoing contest rather than a milestone. That is the point. The strongest evidence a lab can produce still gets tested, and a company that has only its own confidence to offer will not last the first round.
Patent strategy runs on evidence, not on being first
The same test applies to any business that trades on a technical edge. The questions it will face are predictable:
– Does the patent specification actually teach the invention, with the best method the company knew at filing? Both are grounds of invalidity, and both are decided on what was written down at the time.
– Can the data behind the claimed result be reproduced, and is there a record of who did what, when, with which materials and code?
– If the edge is kept as a trade secret, can the company show what the secret is, that it was treated as confidential, and that the departing engineer took it (and, or course defend against lack of Best Method attack)?
– In a sale or a capital raise, will the technical claims in the information memorandum hold when the buyer’s advisers ask for the underlying records?
None of this is generated on the day it is needed. When Your Patent Isn’t Built for the Road Ahead shows an Australian patent failing on insufficiency and best method because the specification could not carry the weight placed on it. Built Before the Fight: What May’s IP Decisions Reward makes the wider point that disputes are decided by records made years earlier. And Where Is Your Leverage? What a Week of IP Disputes Teaches About Strategy argues for stress-testing patents and documenting data provenance before anything is asserted.
The IBM team built verification into the experiment rather than bolting it on afterwards. That is the model. Treat the proof as part of the invention, budget for it, and make someone responsible for it. A lead you cannot prove is a lead you do not own.
Read the article: IBM quantum computer solves classically intractable problem in 15 minutes, ScienceDaily, from materials provided by IBM and the University of Chicago.
Capability is not what separates an attacker from a defender. Incentives are.
Ben Thompson’s Stratechery piece of 24 August 2026 works this through in agentic cybersecurity, and the logic travels a long way past security. An attacker’s automated agent carries positive expected value. If the exploit fails, nothing has changed. If it lands once, the attack has paid for itself. A defender’s automation carries the opposite. A good patch only preserves the status quo, while a bad one breaks the software or opens a fresh hole. So offence automates completely, defence keeps a human in the loop, and no human in the loop keeps pace with an agent. Thompson then applies the same asymmetry to incumbents and startups. An established company judges AI by the mistake it might make, so AI stays a productivity tool. A startup, whose base case is failure already, has nothing to lose by automating everything. Same tools, different incentives, very different outcomes.
What this means for IP strategy
The durable lesson is that this risk calculus is quietly setting your filing decisions as well as your deployment decisions. Three things follow.
– The defender’s advantage is ownable. Only the defender holds the code, the dependency map and the operational telemetry. That advantage compounds only if it is captured – patents over the closed remediation loop of detect, propose, deploy and roll back, and controlled trade secrets over the pipelines and data that make the loop work. It is the same pattern as in When Everyone Can Run the Model: What Open-Weight AI Means for Your IP Strategy – once the capability commoditises, what is left to own is the machinery around it.
– The human you remove from the loop was also doing IP work. They knew what was confidential, what was privileged and where a draft came from. Take them out without rebuilding the controls and the reasonable steps your trade secret protection rests on leave with them, which is the exposure in Your AI Agent Won’t Keep a Secret.
– The incentive gap shows up on the register. Companies treating AI as a productivity gain file nothing new. AI-native entrants file around workflows that did not exist two years ago, which is the ownership question in When Anyone Can Build It in an Afternoon, What’s Left to Own?.
The useful question for your next board paper is not how much AI you have deployed. It is what you filed, and what you locked down. Caution about deploying is defensible. Caution about owning is a slower way to lose the same position.
Read the article: Autonomy and Innovation, Ben Thompson, Stratechery, 24 August 2026.
A rebrand does not just launch a new name. It starts a clock on the old one. Stop using a trade mark, show no intention of going back to it, and the law can treat it as abandoned. At that point a competitor can pick it up, together with the brand equity you spent years building. That is the trade mark abandonment risk most rebranding plans never mention.
A nice example comes from the ruling in the Delaware federal court in the dispute between X Corp and Operation Bluebird, a startup that set out to relaunch the old Twitter brand. The judge granted X a preliminary injunction over the TWITTER name itself. On the word TWEET and the blue bird logo, X lost. The court found the startup likely to prove that X had discontinued genuine use of both and did not intend to resume it. The startup renamed its platform Tweet.app the same day, having already taken more than 172,000 handle requests before launch.
Rebranding without losing your trade marks
The strategic point is that the abandoned assets were not dead assets. Two were valuable enough for a challenger to build a business on. The TWEET mark and the bird still carried more recognition than most companies ever earn, and they were left on the table.
Three things worth taking from it:
– Abandonment is a question of conduct, not intent. A rebrand memo does not preserve the old marks. Continued genuine use does. Where the old brand still has value, keep it in real commercial use somewhere, even at small scale, or decide deliberately to let it go.
– Different marks fail separately. X kept TWITTER because it kept using it. The word, the logo and the product vocabulary each stand or fall on their own record of use, so an audit after a rebrand has to go mark by mark.
– The public does not rebrand on instruction. “Tweet” stayed in everyday speech long after the company dropped it. Where the market keeps using your old name, a competitor adopting it gets recognition for free.
Most countries run the same logic on a fixed timetable. Once a registered trade mark has gone without genuine use for a set period, usually three to five years, anyone can apply to have it removed. A rebrand starts that period running on every mark it retires.
The mistake is treating retired brand assets as sunk cost. Walking Away From a Brand Doesn’t Always Mean You’ve Let It Go set out the challenge when it was filed. It is not a new weakness either. No brand protection – Twitters greatest challenge? flagged the thin protection around the bird and the word “tweet” in 2009. And it is the same pattern as The Quiet Decay: Why IP Value Slips When No One Is Watching – the loss came from inattention, not from a stronger opponent.
Before the next rebrand, list every mark being retired, decide for each whether it is kept in use, licensed or released, and put someone’s name against the decision. Otherwise the market, and eventually a court, will decide for you.
Read the article: Judge blocks X rival from using Twitter name, but allows ‘Tweet’ for now, Sarah Perez, TechCrunch.
Global banking has just posted the most profitable year of any industry – US$1.3 trillion in net income – and still trades at the lowest price-to-book of any industry. McKinsey’s preview of its Global Banking Annual Review 2026 reads that gap plainly: investors are delighted with the results and unconvinced the advantage will last.
The report shows why. Banking’s real competitive moat was never property. It was inertia – trust built over decades, relationships too tangled to unwind, customers who never got around to chasing a better rate. Four forces reached a tipping point in 2025, and every one attacks that inertia directly. Mature fintechs have taken 17 per cent of the revenue they share with the top thousand banks. Neobanks like Nubank (131 million customers) and Revolut (69 million) now beat incumbents on growth, on returns and, increasingly, on trust. Stablecoins loosen the deposit base. And agentic AI dissolves stickiness itself: software agents that monitor balances in real time, sweep idle cash into higher-yield accounts, shift card balances and compare products while the customer sleeps. Banks survived the internet and the smartphone by moving at the pace of their older, high-value customers. That grace period is gone – generative AI reached roughly half of US working-age adults within three years, across every age group at once. When technology can dissolve your switching costs without infringing a single right you hold, what remains is what you actually own.
When inertia dissolves, ownership decides: the intellectual property strategy lesson
The durable lesson is not about banking. Any business whose advantage rests on habit, incumbency or switching costs is holding a moat that technology can drain without ever infringing it, because inertia cannot be registered, licensed or enforced.
Now look at where McKinsey says the surviving value sits – hyperpersonalised engagement built on proprietary customer data, platforms that reach beyond the core product, agentic operations, a three-speed innovation portfolio – and you are reading a list of intangible assets. Each one is protectable: the data and the models trained on it, the operational know-how as trade secrets, the platform technology as patents, and the brand that customers and their AI agents must still choose to trust as registered trade marks. It is the same shift I examined in When AI Does the Shopping, What Does Your Brand Actually Own? – when an AI intermediary sits between you and the customer, owned rights are what travel – and in When Everyone Can Run the Model: What Open-Weight AI Means for Your IP Strategy – when a capability commoditises, value moves to cost position, proprietary data and whoever owns the customer relationship. Banks already spend more on technology than the next four sectors combined; the strategic question is how much of that spend ends the year as owned, enforceable assets rather than vendor licences and undocumented practice – the discipline set out in When AI is embedded in your Workforce, Trade Secrets Become the Strategy. Three questions for your next strategy review, whatever your industry:
– Which parts of your customer relationship would survive the arrival of an AI agent acting for that customer, and which are merely habit?
– What did last year’s technology spend leave behind that you own – filings, controlled trade secrets, data rights – rather than rent?
– If a competitor can buy the same AI you can, what in your stack could they still not copy?
Investors are already pricing the answers. The time to convert record profits into owned advantage is while the profits are still records.
Read the article: Global Banking Annual Review 2026: Precision with speed, McKinsey & Company, 21 May 2026.
You must be logged in to post a comment.