Iran fired ballistic missiles and drones at US bases in Kuwait and Bahrain on Sunday, retaliating for US strikes on five Iranian coastal sites a day earlier. Within hours both sides agreed to halt attacks and meet in Doha on Tuesday. Strategy's STRC preferred sank to record lows as its coupon ratchet accelerated a spiral the instrument was designed to prevent, and GPT-5.6 went live in government-gated limited preview.
The overnight delta is confirmation: the weekend halt held into Monday and the Doha talks are set for Tuesday, the de-escalation path the Geopolitics section maps.
Markets read it as relief rather than risk, with equity futures opening firmer, crude staying contained instead of gapping, and gold and rates near pre-strike levels. The Dashboard carries the levels.
Crypto data provided by CoinGecko
Three weeks of falling oil attracted short energy positioning and long rate-cut proxies into a market that had priced out the war premium entirely, and the weekend just introduced an asymmetric unwind risk. WTI closed Friday near $70, roughly where it was before the strait crisis began. That price embedded a specific assumption: diplomatic progress converts into reopened shipping lanes. The weekend's events tested that assumption with ballistic missiles on allied soil. Monday delivered the answer: WTI held near $70, briefly tapping above before easing once the halt-and-Doha headline landed, so the tape read the escalation as containable rather than supply-destroying. That is the benign branch, but the positioning risk did not clear, it merely went unpunished: the entire energy short book is still built on a peace thesis, and the next headline that breaks the wrong way meets the same one-sided book. The last time oil positioning was this one-sided into a geopolitical weekend was early February, and the gap that followed added $8 in a session.
The Nasdaq's five-session losing streak hardened into something structural when growth catalysts fired and failed to reverse the rotation. The week included OpenAI's biggest product launch and the best biotech catalyst in years (Moderna's 9-0 FDA advisory vote). AI news did not rotate capital back into growth. Biotech drew capital into defensives. When positive sector-specific catalysts cannot reverse a leadership rotation, the rotation is no longer responding to headlines. It is responding to the rate regime. The last time defensive leadership held this firmly while growth catalysts fired and failed was Q3 2022, the quarter before equities gave back three more months of gains into the October low. A rotation that keeps ignoring good news has stopped being a trade and started being a forecast.
Long Lake Management, an AI-native holding company backed by General Catalyst and Alpha Wave, agreed in May to take American Express Global Business Travel (NYSE: GBTG) private for $9.50 a share, roughly $6.3 billion, a 60 percent premium, with American Express, Expedia, the Qatar Investment Authority, and BlackRock committed to vote yes. The buyer's stated playbook: acquire labor-heavy services businesses and rebuild them on AI. Corporate travel is the ideal target: thousands of human agents booking trips, a cost base an agent-grade model can absorb. The bet is a change of input, not cost-cutting. Convert a headcount business into a software business, then keep the margin. This resembles Genpact's 2005 spinout from GE Capital, which industrialized back-office labor by shipping it to Manila. The AI roll-up runs the same arbitrage, but the cheaper worker is the model, not the offshore employee. The counter: a 60 percent premium into a hawkish rate regime is a steep entry, and corporate travel retains accounts through relationships that over-automation erodes.
On June 23, tokenized equities reached 17 percent of Solana's daily decentralized-exchange spot volume and, for the first time, out-traded memecoins as a share of the chain's daily activity. The next day tokenized-stock volume hit a $644 million record, triple the mark set on June 16. For the prior week Solana cleared roughly 95 percent of all on-chain tokenized-stock volume, its tokenized-stock holders crossed 200,000 wallets, and stablecoin supply on the chain passed $16 billion. A blockchain becomes what trades on it, and Solana just crossed the line where real-world equities out-trade the memecoins it was ridiculed for. The holder count signals retail distribution, not a handful of arbitrage desks. NASDAQ launched in 1971 as an automated quotation board for over-the-counter penny stocks, dismissed as a dealer's backwater, and grew into the venue where Apple chose to list. The moment to re-rate a speculation venue is when its composition flips.
Strategy's STRC preferred, roughly $10.5 billion outstanding at an 11.5 percent coupon, hit $74.57 on Friday, a record low that has triggered the contractual coupon ratchet multiple times since launch at $90 and 9 percent last July. Below $95, Strategy must raise the coupon 50 basis points on all outstanding shares, adding approximately $53 million in annual dividend costs. The cost increase deepens the credit doubt, which drives the price lower, which triggers the next ratchet. Grayscale's Zach Pandl and Ripple's Brad Garlinghouse now argue publicly that a $3-billion-plus Bitcoin sale would restore confidence better than a coupon hike. This resembles the floorless "death-spiral" convertibles of 1998 to 2001, whose terms reset against the issuer as the stock fell, accelerating each decline until several spiraled into delisting. The sub-$95 break revealed the trigger was there all along, embedded in the coupon, not the balance sheet. A stability mechanism that depends on the price it is meant to stabilize is an accelerant wearing a brake's label.
GPT-5.6's real news isn't that it launched June 26 government-gated to roughly twenty pre-approved organizations, a gate telegraphed all week, but that OpenAI priced it as three products at once. Sol runs $5/$30 per million input/output tokens, Terra $2.50/$15, and Luna $1/$6. Luna undercuts the economics of self-hosting most open-source alternatives. Sol introduced "ultra mode," which chains sub-agents behind a single API call to decompose complex tasks, the first commercial offering of multi-agent orchestration as a default serving mode rather than a developer-assembled workflow. The structural question is which tier captures more revenue. If Luna dominates, frontier AI becomes a utility and margins compress toward cloud compute. If Sol dominates, the orchestration layer, not the base model, becomes the moat. The tiered launch reveals OpenAI's market read: simultaneously racing to the bottom (commodity inference at $1/M tokens) and to the top (premium orchestration at $30/M tokens), betting the market bifurcates rather than converging on a single price point.
Anthropic is reportedly on track for its first operating profit, which would make it the only frontier lab whose research-to-product pipeline generates margin instead of burning it. Figures surfaced this week: roughly $10.9 billion in second-quarter revenue, a $44 billion annualized run rate, and $559 million in operating profit. The number matters less than what profitability does to the financing model. Google, Meta, and Microsoft's AI divisions are cost centers subsidized by other businesses; OpenAI reported losses of approximately $5 billion in 2024. A profitable Anthropic changes the industry structure: frontier labs no longer require perpetual venture capital if the product funds the next training run. A lab that self-funds operates under different pressures than one on a fundraising clock; it can delay a launch for safety testing without losing investor confidence. If sustained, it shifts the industry's defining question from "who can raise the most?" to "who can sustain the margin?" The first is a pitch deck problem. The second is a product problem.
Iran's Islamic Revolutionary Guard Corps launched ballistic missiles and drones at the US Ali Al Salem Air Base in Kuwait and the US Fifth Fleet headquarters in Bahrain between 2:00 and 3:00 AM local time on Sunday, retaliation for US strikes on five Iranian coastal sites the day before. Kuwait intercepted two missiles with no casualties; Bahrain reported a residential building near the airport damaged. Hitting an air base in a sovereign ally involuntarily conscripts the host nation: Kuwait and Bahrain host US forces on the assumption that the presence deters, not attracts, attacks. Iran inverted that calculus. The near-term question is whether Gulf states demand additional air-defense deployments (deepening the alliance) or quietly explore basing restrictions (loosening it). Historical pattern says the former: after Saddam's 1991 Scud strikes, Riyadh expanded US basing rather than reducing it. But Saudi Arabia's OPEC-plus relationship with Russia and Gulf commercial ties with China give these states exit options the 1991 Saudis did not have.
Within hours of the strikes, the US and Iran agreed to halt attacks and meet in Doha on Tuesday, shifting the venue from Switzerland and narrowing the agenda to the specific question of strait passage. The Joint Maritime Information Center separately widened the navigable route through Hormuz near Oman, a signal that the US is preparing to guarantee passage unilaterally if diplomacy fails. The pattern is Schelling's bargaining theory in real time: both sides demonstrated willingness and capability to hurt, then immediately sat down. The strikes are not a failure of diplomacy. They are the credibility phase that makes the next round possible. Each side now negotiates from proven capability rather than verbal threats. The historical baseline for military demonstrations followed by diplomatic breakthroughs is higher than the reverse: Cuba 1962 produced the Limited Test Ban Treaty in 1963, and "fire and fury" toward Pyongyang in 2017 produced the Singapore summit in 2018. The Doha talks will be more productive because of the strikes, not despite them.
European defense spending commitments face a binding constraint that no budgetary will can solve: production capacity. Rheinmetall, Europe's largest ammunition manufacturer, is expanding from roughly 600,000 155mm shells per year to a target of 1.1 million by 2027, but current consumption in Eastern Europe alone exceeds 2.5 million annually. The factories to absorb defense budget commitments do not exist yet, and manufacturing lead times for armored vehicles, air-defense systems, and precision munitions run 24 to 36 months. The gap between announced spending and deliverable capability is the opening an adversary times its move into, and it does not close by writing checks. It closes by building production lines, which requires the kind of sustained industrial planning that post-Cold War Europe dismantled because NATO made it unnecessary. The same inversion runs through this morning's brief: Strategy's coupon ratchet, built to stabilize STRC, now accelerates its collapse; the US bases meant to deter Iran instead drew its missiles; the safeguard keeps becoming the exposure.
Researchers demonstrated that plague bacteria (Yersinia pestis) were already killing humans 5,500 years ago, long before the cities, grain stores, and rat-infested conditions traditionally cited as prerequisites for plague pandemics. Genomic analysis of ancient remains from multiple Neolithic and Bronze Age sites across Eurasia recovered plague DNA from individuals who lived in small, dispersed agricultural communities. The finding overturns a central assumption of epidemic history: that plague required urban density to become lethal. The pathogen circulated and killed in populations of a few hundred, which means the selective pressure driving virulence evolved in a fundamentally different environment than the crowded medieval cities where plague became famous. Our model of how pandemics emerge may be built on the wrong population density.
An international research team demonstrated that artificial photosynthesis systems can now operate without batteries by automatically adapting their output to match changing sunlight intensity, eliminating the energy-storage bottleneck that has limited every previous solar-to-fuel approach. Prior artificial photosynthesis prototypes produced hydrogen or other fuels at a fixed rate tuned to peak sunlight, which meant they either wasted energy at high intensity or stalled at low intensity. The adaptive system modulates its electrochemical pathway in response to photon flux, maintaining conversion efficiency across the full range of a natural day. The implication is architectural: if solar-to-fuel systems can run without batteries, the cost model for green hydrogen drops by roughly the cost of the battery, which is currently 30 to 40 percent of the system.
A volunteer combing through public radio-telescope data spotted a galaxy moving so fast through space that it has thrown up a glowing arc of plasma nearly 1.8 million light-years long, shaped like a drawn bow and arrow. The galaxy, catalogued as RAD-BAARG, is falling supersonically into a distant cluster, and the friction of that infall against the cluster's thin gas has lit a bow shock, a structure long predicted but almost never seen at this scale. The discovery came not from a professional but from a citizen scientist with India's RAD@home collaboratory, reviewing ultra-sensitive images from the LOFAR radio array. Two truths sit in one object: the sky still hides features the size of galaxies, and the eyes that catch them no longer need a doctorate.
A review of how unrelated species cooperate found that the coordination runs on real, evolved communication, not instinct or coincidence: the greater honeyguide, a wild African bird, leads human foragers to bees' nests and adjusts its calls in response to theirs, splitting the wax-and-honey reward. Drawing on birds, fish, insects, and mammals, fifty-eight researchers writing in Animal Behaviour show that cross-species signals are flexible, sometimes learned and regionally varied, and built to solve a hard problem: telling a trustworthy partner from one about to cheat. Cleaner fish advertise their role with bright colors and a distinctive dance so predators hold still instead of eating them. The lesson travels: durable cooperation between parties with nothing in common rests less on goodwill than on a shared, hard-to-fake signal.
Food and the battery boom are about to bid against each other for the same molecule, and the refinery chokepoint that feeds both sits inside the one country now fencing off its exports.
Phosphate has spent a century as a sleepy fertilizer input, and it is quietly becoming a contested strategic mineral. Lithium-iron-phosphate (LFP) cells are now roughly 47% of the global battery market, and McKinsey expects them to hold that share through 2030, which means the energy transition now needs battery-grade purified phosphoric acid, an estimated 2.5 million additional tonnes by 2030, with cathode demand for the material rising more than twentyfold against 2020. The collision is not a 2030 story; supply shocks are pulling it forward into this year. A worsening sulphuric-acid shortage, and about 40% of sulphuric acid goes into making phosphoric acid, is throttling DAP and MAP fertilizer output into 2026-27; China's NDRC has suspended phosphate-fertilizer exports through August 2026 and widened the curbs to more blends, so between half and three-quarters of Chinese fertilizer exports are now restricted; and China, which refines more than 95% of the world's purified phosphoric acid, has signaled it will halt sulphuric-acid exports starting May 2026. So the same country that controls battery-grade phosphate refining is simultaneously walling in its fertilizer phosphate, and phosphate prices are expected to stay elevated through at least the end of 2026. The structural point the market hasn't priced: phosphate has become a dual-use input, food security on one side, the battery supply chain on the other, running through a single nation's export policy, and the West's "cobalt-free, nickel-free" LFP story has a hidden Chinese-phosphate dependency underneath it. If DAP/MAP prices firm into early 2026 as CRU expects while battery-grade purified phosphoric acid stays overwhelmingly Chinese, expect a durable re-rating of the few Western owners of integrated, non-China phosphate, Mosaic (MOS), Nutrien (NTR), and ICL Group (ICL), whose rock and acid turn strategically scarce on two fronts at once, while the squeeze lands on whoever cannot substitute: farmers eating higher fertilizer bills, and the Western battery and auto makers whose "diversified" LFP cathodes still trace back to Chinese acid. In plain terms: one Chinese export desk now sets the price of both your groceries' fertilizer and the "cheap" battery, two markets that never knew they were competing for the same molecule, and the few non-China owners of integrated phosphate are the only players standing outside the chokepoint. Watch: the DAP/MAP price (CRU/Argus) and China's phosphate and sulphuric-acid export quotas around the August 2026 NDRC review, plus any Western battery-grade PPA purification capacity that actually breaks ground. If DAP/MAP holds firm through the China export window while no material non-China acid capacity comes online, the food-versus-battery phosphate squeeze is structural, not a one-season fertilizer spike.
The highest-margin business in the grocery store is the ad next to the cereal, and Google's quiet decision to keep the cookie just stripped the open web of its last excuse to catch up.
Retailers are turning their checkout data into the most profitable business they own, and the market still values them as if they only sell groceries and gadgets. US retail-media advertising reaches roughly $71 billion in 2026, about 90% of it captured by two players: Amazon Ads ran $56.2 billion globally in FY2024, Walmart Connect did about $6.4 billion in 2025 (up roughly 46% year over year), and Instacart and DoorDash each clear close to $1 billion. The number that reframes it is the margin: this is 70-to-90-percent-gross-margin ad revenue landing on a retail base that earns a low-single-digit operating margin, so advertising is becoming a disproportionate share of these companies' profit growth even though they are still multiplied as a grocer and a marketplace. The fresh catalyst is the part the consensus has backwards. The standard story held that retail media's tailwind was the death of the third-party cookie, but in October 2025 Google shut down its Privacy Sandbox and kept third-party cookies in Chrome with no removal date, which "should" have rescued the open web. Instead it clarifies the real driver: advertisers don't move budget to Amazon and Walmart because the open web went dark, they move because a closed loop can prove a sale happened with deterministic purchase data. Cookies surviving doesn't save the open web; it just removes its excuse, because the migration was always about measurement and provable return, not privacy plumbing, and that makes it structural rather than a one-off scare. If retail-media revenue keeps compounding north of 20% while Walmart, Kroger (KR), and Costco (COST) widen their ad units, expect the market to slowly re-rate the data-owning retailers, Amazon (AMZN), Walmart (WMT), and Instacart (CART), toward their high-margin advertising earnings, while value drains from the businesses that sell audience inference instead of purchase outcomes: the open programmatic web and legacy retargeters like Criteo (CRTO), plus the long tail of ad-supported publishers fighting over the dollars retailers now keep for themselves. Plainly: the company that owns the checkout, not the one that owns the eyeballs, captures the ad dollar, and the market has not finished repricing who that is. Watch: the retail-media segment disclosures inside Walmart's and Amazon's quarterly results, and eMarketer's retail-media share of total US digital ad spend. If retail media pushes past roughly $75 billion and a fifth-plus of US digital ad spend in 2026 while open-web programmatic CPMs stay soft, the migration to purchase-data owners is structural, not cyclical.
The two Signals share no mechanism, which is the point. The first is a collision: two unrelated demand curves, food and the energy transition, converging on one constrained molecule until the owner of the molecule is repriced by a scarcity it never created. The second is a migration: value sliding from whoever owns the eyeballs to whoever owns the checkout, on no scarcity at all, just a better proof of sale. Different physics entirely: one is a thing too concentrated to share, the other a margin too good to stay hidden. Track the input in the first and the profit mix in the second, and in both the reported number arrives about a year after the structure already decided it.
The Fungible Bottleneck: when one constrained production line feeds both a premium and a commodity market, the two end-products become substitutes in production, so a demand surge in the high-margin one is a supply shock to the low-margin one, and the cost resurfaces in a market that looks unrelated to the buyer who caused it.
Apple, the most powerful hardware buyer on earth, is petitioning Washington for permission to buy memory from CXMT, a Chinese firm on the Pentagon's PLA-linked blacklist. The reason: AI's hunger for high-bandwidth memory (HBM, the stacked DRAM that feeds AI accelerators) has made ordinary phone-and-laptop memory so scarce that Apple has already raised Mac and iPad prices 17 to 25 percent, with Tim Cook calling it a "hundred-year flood." Micron has abandoned the consumer-memory business entirely.
What surface analysis misses is that the AI-bubble debate asks only one question: will the roughly $750 billion of 2026 infrastructure spend pay off for the hyperscalers who spent it? That framing hides the fact that much of the bill is already being paid, by people who never bought a GPU. HBM earns three to five times more revenue per wafer than commodity DRAM, so every wafer the three firms that control 95 percent of the market move toward AI is a wafer denied to phones, cars, and consoles. Data centers now absorb roughly half the world's DRAM, up from a third five years ago; commodity contract prices nearly doubled last quarter. The money didn't vanish into "capex"; it crossed the bottleneck and reappeared as a regressive consumer tax that no inflation line attributes to artificial intelligence.
The call: through 2026, memory becomes a named, recurring driver of consumer-hardware prices and a measurable contributor to goods inflation, even as consensus insists goods are disinflating. Smartphone and PC unit sales contract, by IDC's roughly 5 and 9 percent, on price-driven demand destruction, and at least one major maker beyond Apple raises prices explicitly citing memory before year-end. The "AI capex is the hyperscalers' problem" trade keeps misfiring because the first balance sheet the buildout impairs is the household's.
Where this might be wrong. The strongest objection is that this is a cycle, not a regime. Memory is the most violently mean-reverting market in technology; the cure for high prices is high prices. At these margins, Samsung, SK Hynix, and Micron have every incentive to build, and history punishes permanence: the 2017-18 super-cycle wore the identical "structural shortage" story and ended in a glut that round-tripped every gain, taking DRAM down more than 60 percent. If greenfield fabs and node migration relieve the crowd-out by mid-2027, the consumer AI tax was a spike mistaken for a transfer. The second failure runs the opposite way: if data-center memory demand is itself the bubble the BIS just warned about, an AI pullback would free commodity capacity and flip the tax into a windfall, the same bottleneck unbinding in reverse. And the household incidence may be smaller than the headline: memory is a modest share of a phone's bill of materials, so even a doubling lifts the shelf price by single digits, real but short of regime-defining. The framework fails if contract DRAM and NAND prices go flat-to-down quarter-over-quarter by Q4 2026, or if smartphone and PC units stop shrinking; either would prove the shared line no longer binds.
"Whatever inspiration is, it's born from a continuous I don't know."
— Wisława Szymborska, Nobel Lecture (1996)
Szymborska was a Polish poet who spent most of her life in Kraków, won the Nobel Prize in Literature, and was known for being unable to explain what poetry was. Not because she lacked vocabulary. Because she believed the honest answer was that she did not know, and that the not-knowing was where the work came from. Her Nobel lecture was about the phrase "I don't know." She argued it was the most important sentence in any language because it was the one that preceded every real question, every real discovery, every real change of direction. She thought the people most dangerous to understanding were the ones who had stopped saying it.
There is a version of this that costs you something every day. It is the question you keep answering with borrowed confidence because the admission would cost you standing. The forecast you present as analysis, the skill you list as proficient, the opinion you defend because reversing it feels like losing. The cost of performed certainty is not that you are wrong. It is that it seals you off from the information that would make you actually right. You stop asking because asking reveals the gap. The gap widens precisely because you stopped asking. The paradox is tight: the person who admits not-knowing learns faster than the person who performs knowing, and over any time horizon longer than a single conversation, the learner wins. But in any single conversation, the performer looks more competent. So the incentive is always to perform, and the cumulative cost is always paid later, in the form of a decision made on a foundation you stopped testing.
Today's practice: pick the one question you've been answering with certainty that you honestly don't know. The next time it comes up, say "I don't know" and watch what happens. If the admission costs you nothing, the certainty was armor you didn't need. If it costs you something, you just found where your credibility depends on a bluff, and that is the most valuable thing you can learn today.
A technology startup spends three years building its culture around speed: ship fast, iterate loudly, celebrate the quick release over the careful one. By year four, it has a product with serious reliability problems. Leadership decides to pivot to engineering quality. They announce new standards, hire a VP of Engineering from a company famous for its infrastructure, and write a culture document about craftsmanship. A year later, the reliability problems persist and the VP has left. The team kept building fast. The incentive signals, the promotion criteria, the muscle memory of a thousand daily decisions all ran in the grooves worn by years of speed. The culture document described a different valley. The organization was already deep in its own.
C. H. Waddington, a developmental biologist working in the 1950s, developed an image to explain how embryonic cells commit to becoming a particular cell type, nerve or muscle or skin, and then resist being redirected into anything else. He drew a ball rolling across a landscape of ridges and valleys. Each valley represents a developmental pathway, and once a cell rolls into a valley it is pulled by gravity toward that path's endpoint. Moving a cell from one valley to another requires enough force to crest the separating ridge, and as the cell descends deeper, that ridge grows taller. Waddington called this the epigenetic landscape. The mechanism is that each developmental choice reshapes the landscape itself, steepening the walls of the current valley and raising the ridges between it and alternatives.
The decision tool applies in any system that develops through sequential commitments. Before a major change, map the terrain: how entrenched is the current path, and how much disruption is required to escape it? Low entrenchment means ordinary management can shift direction. Deep entrenchment means the organization requires enough destabilizing force to escape the groove before it can settle into a new one, which typically means tolerating a period of worse performance before improvement arrives. The failure mode is treating a high-entrenchment problem with low-entrenchment tools: a culture document, a new hire, a reorganization. Those interventions work on flat ground. Against a steep gradient, they dissipate without moving the system. The diagnostic: if a change initiative is in its second year and behavior looks identical, you are not applying the wrong solution. You are applying insufficient force to an obstacle you have not measured.
Mix crystals of different sizes into a solution, or scatter grains through an alloy, leave the whole thing alone, and something quietly violent happens: the small ones vanish and the large ones grow. Around 1900 the chemist Wilhelm Ostwald described why, and the process now carries his name. Ostwald ripening. The driver is not competition; the particles never touch, never sense each other, never fight. It is surface energy. A small particle has more surface curving around less volume, and that tight curvature raises the chemical potential of the atoms on its skin, the Gibbs-Thomson effect, making a small particle very slightly more soluble than a large one sitting in the same liquid. So atoms drift off the small particles, diffuse through the shared medium, and redeposit onto the large ones, all of it rolling downhill toward one objective the system never announces: minimize the total surface area. The Lifshitz-Slyozov-Wagner theory worked out the arithmetic in 1961 and found the ending is universal. Fewer, larger particles, arriving on a predictable schedule. No particle does anything to any other. They are coupled only through the solution they all happen to share.
The reflex, watching small players steadily disappear while big ones swell, is to reach for a story of combat: the large are out-executing the small, undercutting them, eating them. Ostwald ripening says inspect the medium before you write the rivalry. Some of the most relentless consolidation in the world involves no rivalry at all. The units never interact; they are merely suspended in a common field (a labor pool, a capital market, an attention economy, a regulatory regime) that levies a tax on smallness as such, and that tax alone bleeds the small into the large. The signature to look for is the Gibbs-Thomson one: the disadvantage scales with how small you are, not with any head-to-head contest, and it grinds away even on players who will never meet. This is not the familiar "rich get richer" of networks, where nodes actively attract more links; here nothing attracts anything. The large simply offer a lower-energy place for loose material to land.
So when you see consolidation and feel the pull to explain it as the strong devouring the weak, run the Ostwald test before you act on the story: is anyone actually competing, or are the units just sharing a medium that penalizes small size? Try to name the specific rival capturing the share. If you can't name one, you are not losing a fight, you are being ripened, and the moves that work are entirely different. Polishing the small unit's quality is wasted motion, because the gradient cannot read merit; the only real escapes are to get large enough to sit on the receiving end of the flow, or to lift the unit out of the shared field altogether, its own capital, its own channel, its own pool. You will know within a week whether you were honest with yourself: pick one consolidation you are inside or betting on, try to name the competitor taking the share, and if you come up empty, stop optimizing the product and start changing the structural exposure.
(Ostwald ripening. Wilhelm Ostwald, c. 1900; coarsening kinetics formalized by Lifshitz, Slyozov & Wagner, 1961.)