Corporate profit margins broke a 50-year ceiling to reach 20.6% while consumer sentiment sits at record lows and food insecurity is rising. The Iran ceasefire extension remains unsigned as both sides publicly dispute terms. Blue Origin's only launch pad was destroyed in a static fire explosion, stranding Amazon's satellite constellation weeks before SpaceX's IPO.
Crypto data provided by CoinGecko
US corporate profit margins just hit 20.6%, shattering a ceiling that held for fifty years, and the market's inability to explain why is the most important unresolved question in macro. Eric Basmajian at EPB Research flagged the data Friday. Two explanations compete with opposite investment theses. First: AI-driven efficiency gains are permanently lowering operating costs faster than prices, producing real productivity GDP statistics have not yet learned to measure. Second: market concentration has created oligopolistic extraction where a shrinking number of firms capture an expanding share of revenue. The K-shaped data points toward the darker reading. Stocks sit at record highs while consumer sentiment sits at record lows. The savings rate at 2.6%, the lowest since June 2022, means consumers are drawing down buffers to maintain spending. If margins are from efficiency, consumer well-being should eventually follow. If margins are from extraction, they revert violently when the consumer base erodes.
Treasury Secretary Bessent announced that incoming Fed Chair Kevin Warsh will eliminate forward guidance entirely, a structural break from fifteen years of the Fed's primary communication tool. Bessent said Warsh will reduce the number of governors giving speeches, calling it "too many Fed governors popping off." Nick Timiraos captured a second observation that may matter more: the group most committed to crushing inflation in the abstract is the same group least willing to keep rates high to achieve it. If Warsh follows through, every FOMC meeting becomes genuinely live for surprise in either direction. The fixed-income market has not traded without forward guidance since before 2008. An entire generation of bond traders has never operated in an environment where the Fed does not telegraph its intentions. Hedging costs increase, and the carry trade structures built on predictable rate paths lose their foundational assumption.
Blue Origin's New Glenn rocket exploded during a static fire test at Launch Complex 36 on May 28, destroying the only pad capable of launching the vehicle and releasing energy equivalent to 20% of the Hiroshima bomb. The timeline is the story. Amazon needs 1,618 Kuiper satellites in orbit by July 30 to meet its FCC license requirement and currently has roughly 300. New Glenn was the primary launch vehicle. The pad will require months of repairs. SpaceX's IPO is scheduled for June 12 at a $1.75 trillion valuation built partly on Starlink's 10.3 million subscribers. Every month of Kuiper delay is a month where Starlink locks in more subscribers and institutional contracts. If Blue Origin's pad is not operational by Q4, Amazon faces a choice between paying SpaceX for Kuiper launches, an extraordinary competitive concession, or missing the FCC deadline.
Tokenized fund assets under management on Ethereum reached $19.3 billion, up from near zero two years ago, and every major institutional participant chose the same blockchain. BlackRock's BUIDL, Fidelity's FDIT, J.P. Morgan's MONY, and BlackRock's newly filed OnChain Shares all settled on Ethereum. Joseph Chalom at BlackRock wrote that "Ethereum is going back on offense." SoFi launched SoFiUSD, the first stablecoin issued by a nationally chartered bank on a public blockchain. Ethereum's stablecoin supply stands at $179.2 billion, 58% of the $308.6 billion global total. The venue question that dominated crypto infrastructure debates for five years has been answered empirically. The parallel is NYSE versus Nasdaq in the 1990s: reputation, liquidity, and network effects concentrated institutional venue selection then, and the same forces are concentrating institutional crypto activity on Ethereum now.
Jamie Dimon told an audience that Coinbase CEO Brian Armstrong is "full of shit" regarding the CLARITY Act, while Senator Lummis warned this Congress "may be the last crypto legislation window before 2030." Dimon's hostility is the signal. The CEO of the largest US bank is publicly attacking the legislative framework that would formalize crypto as a regulated asset class, while his own institution's digital assets team builds products to serve that same framework. When the incumbent's CEO calls the insurgent's legislative effort illegitimate while the incumbent's product team builds for the world that legislation creates, the institution has already conceded the structural question and is fighting over terms. The CFTC approved the first regulated Bitcoin perpetual futures contract on May 29, directly competing with Hyperliquid and Binance's offshore operations. If the CLARITY Act passes alongside regulated perps, the US recaptures derivatives flow that has been hemorrhaging offshore since 2021.
Meta released ESMFold2 with an atlas of 6.8 billion proteins and 1.1 billion predicted structures, achieving state-of-the-art performance on antibody-antigen interactions without requiring multiple sequence alignments. The model predicts a 1,024-residue protein in nine seconds. AlphaFold3 covers 200 million structures. ESMFold2's 1.1 billion represents a five-fold expansion of the structural universe available for drug design, and it is fully open-source. Lab validation showed 65% of designed antibody-antigen binders passed functional benchmarks with binding affinities consistent with therapeutic activity. The structural shift is the economics: if protein design follows AI text generation's cost curve, the $2,000 computational screen becomes a $2 query, and the bottleneck in drug discovery moves from finding candidate molecules to running clinical trials on candidates AI generates faster than pharma can absorb.
One in three AI teams ran an open-weights model in April 2026, up from one in five nine months ago, with the total number of teams using open weights tripling, according to LangChain's infrastructure data. Georgi Gerganov launched llama.app, a single-line cross-platform installer for local AI inference. A separate project, pibot, demonstrated a fully local voice AI stack running entirely on-device with zero Python dependencies via Rust and mlx-c. Larry Ellison added the strategic frame: "AI is rapidly commoditizing because most models are trained on the same public internet data. The real competitive edge isn't the model anymore. It's access to exclusive, proprietary datasets." If Ellison is right, the inference layer commoditizes and the business model of selling API access to frontier models faces structural headwinds as open-weights reaches consumer-grade simplicity.
Andover, New Jersey cancelled a data center project and passed a complete ban on data center construction, joining a pattern of $41.7 billion in data center projects cancelled or paused in Q1 2026 as community opposition becomes a political force. The pattern is specific to communities with residential bases and limited industrial zoning: data centers consume enormous power, generate minimal local employment (typically 30-50 permanent jobs), and produce waste heat and noise. The AI infrastructure buildout requires physical space in locations with cheap power and fiber connectivity, which increasingly overlaps with communities that do not want them. The constraint is not silicon. It is permission: the political and physical infrastructure needed to house, power, and cool the chips. If local opposition continues spreading through state legislatures, the bottleneck shifts from chip supply to site availability, and companies with existing permitted capacity gain structural pricing power.
US Defense Secretary Hegseth announced at the Shangri-La Dialogue that India will support and maintain US Navy vessels on forward operations in the Indo-Pacific, a framework India accepted while remaining a founding member of BRICS and continuing to purchase Russian oil. Santiago Capital called it "the BRICS with US Navy characteristics." India is not joining a US alliance. It is providing military infrastructure support to the US while maintaining economic relationships with US adversaries, a posture impossible before the Hormuz crisis demonstrated that security and trade architecture can run on separate tracks. The Indo-Pacific security architecture is being rewired in real-time, with Vietnam stressing trust erosion, South Korea pivoting to self-reliant defense, and India threading a needle between US military cooperation and BRICS membership no country has successfully maintained long-term.
China's construction sector output accelerated its decline to negative 8% year-over-year in Q1 2026 while Beijing launched its most aggressive offshore wealth crackdown in decades, targeting residents with deposits above $30 million. Edward Dowd flagged the construction data as the transition from chronic decline to acute crisis. Michael Pettis identified three paths to raise consumption's share of GDP, all painful: redistribute within households, transfer from businesses to households by raising rates, or transfer from government to households. Pettis: "It is much better for China to begin the difficult adjustment while it still has fiscal space." The offshore wealth crackdown, with penalties on Futu and Tiger and threats to $250 billion in Hong Kong assets, is consistent with capital controls during structural adjustment. Trapping capital inside borders during adjustment historically accelerates the crisis it is trying to prevent.
Russia's oil tax revenues surged from $4 billion in March to $10 billion in April as the Urals crude discount vanished entirely, funded by the same energy dynamics the US-Iran confrontation was meant to resolve. Robin Brooks flagged the data: Western sanctions created the Urals discount to impose costs on Russia, but Iran's Hormuz disruption tightened global supply enough to eliminate the price differential. Every barrel that does not transit Hormuz makes Russian crude more valuable. The paradox is structural: resolving the Iran crisis by reopening Hormuz would help oil consumers but simultaneously cut Russia's windfall. Failing to resolve it sustains consumer pain while funding the war in Ukraine. The energy system is running a feedback loop where one conflict's disruption finances another conflict's prosecution.
Archaeologists discovered six previously unknown Bronze Age mines in southwestern Spain whose copper and lead isotope signatures match metals found in Scandinavian artifacts more than 3,000 years old. The University of Gothenburg team found the mines near Cabeza del Buey in Extremadura, including one site packed with around 80 stone axes used to crush ore. Combined with approximately 20 additional prehistoric mines documented between 2024 and 2026, the discoveries suggest metals from the Iberian Peninsula were traded across the full length of Europe before the Iron Age, a continental supply chain operating without written language, currency, or centralized authority. When the physical evidence contradicts the model of what was possible, the model's assumptions about what coordination requires are wrong.
Physicists at Cal Poly published findings in Physical Review B demonstrating that time-dependent magnetic fields can generate quantum states that have no counterpart in any static material, matter that literally cannot exist unless the system is being driven. The research showed that periodically changing a magnetic field reorganizes quantum systems into topological phases with no equilibrium equivalent. The states created are more stable and error-resistant than their static counterparts. The deeper implication: useful quantum properties depend not just on what a material is but on how it is driven in time. Some things only exist in motion.
Planetary scientists calculated that moons orbiting rogue planets, worlds wandering through the galaxy without a star, could remain warm enough for liquid water through tidal heating and hydrogen-rich atmospheres alone. The finding overturns the assumption that life requires a star. The number of rogue planets in the Milky Way is estimated at billions, potentially exceeding the number of stars. If even a fraction host moons with tidal heating, the habitable real estate in the galaxy is orders of magnitude larger than the "habitable zone" around stars that has framed every previous search for life.
Researchers at the University of Miami's Rosenstiel School discovered that nitrogen cycling in oxygen-depleted ocean zones is far more dynamic than any previous model assumed, with key intermediate compounds appearing and disappearing on timescales that existing monitoring systems were too slow to capture. The old models treated these zones as chemically stable between measurements. The new data shows they fluctuate continuously, with microbial communities adjusting their metabolic pathways in response to conditions that change faster than quarterly sampling can detect. The error is familiar: measuring a dynamic system at static intervals and concluding it is stable because the snapshots look the same.
Vizhinjam's million-TEU milestone and JNPT's sevenfold transshipment surge are locking feeder-network reallocations into multi-year contractual commitments that preclude Gulf-hub reversion.
More than 34,000 shipping route diversions have been recorded since the Hormuz disruption began. India's Jawaharlal Nehru Port in Navi Mumbai has seen transshipment volumes surge 700% above pre-crisis baselines, with average import dwell times doubling from 12 to 23 days. Vizhinjam, India's first dedicated deep-water transshipment terminal in Kerala, hit one million TEU faster than any port in the Indian Ocean's history. India's government has accelerated construction on two additional mega-terminals, Vadhavan and Galathea Bay, specifically to capture the rerouted traffic permanently. The structural logic: once shipping lines reconfigure feeder networks, retrain crews on new routes, and sign multi-year terminal handling agreements, the switching costs to revert make the old routing uneconomical even after Hormuz clears. Dubai's Jebel Ali, which built the UAE's entire services economy on its transshipment monopoly, faces the first credible structural competitor in decades. If Indian transshipment volumes at Vizhinjam and JNPT sustain above 40% of crisis-peak levels for two quarters after Hormuz reopens, the Gulf-centered shipping architecture that has dominated Indian Ocean trade since the 1980s has permanently fractured, and Adani Ports captures structural market share that does not revert.
Watch: India's Ministry of Ports, Shipping and Waterways quarterly throughput data (next release July 2026). Maersk and MSC route schedule announcements post-Hormuz reopening. If either carrier maintains India-hub feeder networks rather than reverting to Gulf transshipment, the restructuring is confirmed by the companies that move 30% of global container trade. Track Jebel Ali monthly TEU figures for the inverse signal.
Net inbound migration turned negative in late 2025, eliminating the only replenishment channel for residential tradespeople whose median age is 42 and whose vocational enrollment has flatlined.
Net migration to the United States turned negative in late 2025 for the first time in over fifty years, removing the single largest source of new construction workers at the precise moment the industry needs them most. Housing starts fell to 1.21 million annualized in April, the lowest since June 2020, and the decline is not a demand problem. Permit applications remain elevated. Builder backlogs are full. The constraint is labor: construction sector job openings have exceeded 400,000 for six consecutive months while domestic pipeline of new entrants has stalled, with vocational training enrollment flat since 2019 and the median construction worker now 42 years old. The mortgage rate lock-in effect compounds the shortage: 80% of outstanding mortgages carry rates below 6%, suppressing existing-home inventory and forcing buyers into new construction, exactly the segment where labor constraints bite hardest. Rate cuts, the market's assumed solution, would unlock some locked-in sellers but simultaneously stimulate new construction demand the labor force cannot meet. If housing starts remain below 1.3 million through Q3 while construction sector job openings stay above 350,000, the housing supply deficit widens through 2027 regardless of what the Fed does, and the shelter component of CPI stays elevated because insufficient supply keeps rents and home prices structurally above pre-pandemic trends.
Watch: Census Bureau monthly housing starts report (next release June 17). BLS JOLTS data for construction sector openings (monthly). Freddie Mac's existing-home inventory tracker. If starts stay below 1.25M for three consecutive months while openings remain above 350K, the labor constraint is binding. Secondary signal: track H-2B visa issuance data.
Dark Output: Why Record Margins and Unremarkable GDP Are Not a Contradiction
Dark Output (economic measurement failure where AI creates enormous productive activity that national accounts cannot capture because the activity either never existed before, does not transact through markets, or is classified to the wrong sector, creating a systematic gap between what the economy produces and what the economy measures) explains why corporate margins broke a fifty-year ceiling while GDP sits at 1.6% and consumer sentiment is at record lows, and why the most important question in macro right now is whether this gap reflects genuine invisible productivity or circular financial engineering.
SemiAnalysis coined the term in a May 29 analysis titled "AI Dark Output: The Visible Cost of Invisible Output." The core argument: AI is creating massive amounts of economic activity that GDP literally cannot see. When a literature review falls from $2,000 to $2, we do not do the same number and pocket the savings. We do them before every project. The activity that AI makes cheap enough to perform routinely was too expensive to perform at all before. This creates entirely new categories of work that have no predecessor in national accounts. There is no line item for "things we started doing because AI made them free." Anthropic's Economic Index (March 2026) found that 37% of AI tokens are used in computers and mathematics, yet the contribution to GDP from software investment has not broken from its pre-AI trend. SemiAnalysis calls this "sector misrouting": a hospital uses AI to process paperwork, but the economic value appears in the AI vendor's revenue while the hospital's sector looks stagnant. The accounts count the screws while missing the houses being built with them.
The framework connects three data points that otherwise seem contradictory. First, economy-wide profit margins at 20.6%, shattering a ceiling that held through every cycle since 1975. If margins are expanding but GDP is not, something is producing value that margins can capture but GDP cannot measure. Dark Output fills this gap: AI-driven efficiency reduces costs without creating new market transactions, so margins widen while measured economic activity stays flat. Second, EPB Research's finding that only 1 of 3 recessionary components (residential investment, durables consumption, business equipment) is contracting. The economy is not in recession by traditional measures, yet it feels recessionary to consumers whose real disposable income fell 1.4% year-over-year. Dark Output explains the disconnect: the productivity gains accrue to firms (wider margins) not workers (flat or declining real wages). Third, the clustering of bubble calls from sophisticated allocators. BCA Research's Peter Berezin argues "the AI bubble is primarily an earnings bubble rather than a valuation bubble." Lux Capital's Josh Wolfe holds dry powder as his largest position. Goldman reports record call volume. These allocators are not wrong that something is off. They may be wrong about what it is. If Dark Output is real and large, the margins are earned, not circular, and the "bubble" is actually a permanent regime shift in corporate profitability that traditional valuation models misattribute to overvaluation.
The investment implication is specific. If Dark Output is the primary driver of the margin expansion, the sectors with the highest AI adoption intensity should show the widest margin improvement. Software, financial services, and professional services would lead. Energy, materials, and utilities would lag. The margin dispersion by sector becomes the diagnostic: concentrated AI-sector margin expansion supports the Dark Output thesis. Broad-based margin expansion across sectors with minimal AI exposure suggests an alternative cause (pricing power, financial engineering, labor suppression).
Where this might be wrong, and why Dark Output may be the wrong framework for what is actually happening. The strongest objection is that margins at 20.6% with declining real consumer income is more consistent with extraction than creation. If AI were genuinely producing new economic value at scale, you would expect consumer well-being to eventually improve, either through lower prices or higher real wages as productivity gains diffuse. Instead, food insecurity is rising, the savings rate is at 2.6%, and consumer sentiment is at record lows. The K-shaped data suggests the margin expansion comes from market concentration and pricing power, the same forces that have widened margins in every oligopolistic industry for decades, dressed up in AI narrative. "Dark Output" as a concept contains a circularity: it asserts that AI's impact is real but unmeasurable, which is epistemologically indistinguishable from asserting that AI's impact is small but narratively inflated. The null hypothesis is simpler: maybe AI is not yet producing much transformative output, margins are high for old-fashioned reasons (low borrowing costs for the past decade, labor market suppression, regulatory capture), and the "measurement problem" is a convenient explanation for why the promised AI revolution has not shown up in the data.
Berezin's "earnings bubble" framing offers a second counter. Tintin Capital showed that 58% of Google's Q1 net profit and 52% of Amazon's came from unrealized mark-to-market gains on AI company stakes (Anthropic, CoreWeave), not operating earnings. Under FASB ASC 321, each new AI funding round at higher valuations mechanically inflates the reported earnings of every public company holding stakes, creating a reflexive loop where higher AI valuations improve mega-cap earnings, which support index prices, which validate the AI narrative, which attract more capital at higher valuations. No additional dollar of operating income is required. If the margin expansion is partially fueled by mark-to-market accounting rather than operating efficiency, then it reverses the moment a single major AI company raises at a flat or lower valuation. The margins are real. The question is whether they are earned or engineered.
The test: Two diagnostics converge in Q3. First, the BEA's Q2 GDP advance estimate (late July) alongside the June 25 Q1 third estimate. If GDP prints above 2.0% while margins hold above 20%, the economy is producing measurable growth and the margin expansion is conventional. If GDP stays below 1.8% while margins stay above 20%, the gap is real and Dark Output is the leading candidate for explanation. Second, Q2 earnings season (July-August): examine margin improvement by sector. If AI-intensive sectors (software, financials, professional services) show margin expansion two to three times broader market levels while non-AI sectors show flat or contracting margins, Dark Output is concentrated in AI-exposed firms. If margin expansion is broad-based across sectors with minimal AI exposure, the cause is structural market power, not AI productivity, and the Dark Output framework overestimates AI's current contribution to the economy.
You have been adding. The new app, the extra check-in, the additional step in the process. You have been acquiring because acquisition feels like progress and progress feels like safety. But somewhere you crossed a line you did not notice, where each addition began subtracting from the thing it was meant to serve. The morning routine that has become so elaborate it produces the exact anxiety it was designed to prevent. The productivity system that requires more maintenance than the work it organizes. The research phase that has become indistinguishable from the avoidance it replaced.
"Everyone knows the usefulness of what is useful, but no one seems to know the usefulness of what is useless."
— Zhuangzi, Chapter 4
Zhuangzi tells the story of a carpenter who passes an enormous, ancient tree at a village shrine. His apprentice asks why they do not cut it for timber. The carpenter dismisses it: the wood is knotted, twisted, unsuitable for construction. That night the tree appears in his dream and says: all the straight, useful trees were cut down young. My uselessness is why I am still alive. What you call a defect is my survival strategy. The insight is not that usefulness is bad. It is that a life organized entirely around usefulness has eliminated the things that cannot be optimized but that make the whole structure resilient. The relationship that exists for no reason other than enjoyment. The skill you practice with no intention of monetizing. The afternoon you protected from productivity not as laziness but as a dimension of yourself your metrics cannot see and therefore cannot value.
Identify one thing in your life that you have kept despite its apparent uselessness, something that contributes nothing measurable but that you would miss if it disappeared. Protect it today. Not by justifying it. By recognizing that the inability to justify it might be the point.
The largest terrestrial predator that ever lived was Spinosaurus aegyptiacus, fifteen meters long, dominant across North Africa 95 million years ago. It ate everything smaller. Nothing could challenge it. Then the Cenomanian-Turonian boundary event shifted the climate, altered the river systems it depended on, and Spinosaurus vanished without descendants. Its competitors, smaller theropods with lower caloric requirements and faster reproductive cycles, survived and diversified into the ecological space the giant had occupied.
Kodak employed 145,300 people at its peak in 1988, controlled 85% of the US film market, and generated $16 billion in revenue. Each year it grew larger to consolidate its position. When digital photography arrived, Kodak's size was not a buffer but a burial ground: fixed costs, manufacturing infrastructure, and organizational inertia all scaled with the body the company had built to dominate film. Fujifilm, roughly one-third Kodak's size, survived by pivoting to cosmetics and medical imaging. Kodak filed for bankruptcy in 2012.
Edward Drinker Cope observed in 1896 that animal lineages exhibit a persistent trend toward larger body size over evolutionary time. Each generation that grows slightly larger gains competitive advantages: better access to resources, greater defense against predators, dominance in mate selection. The trend is rational at every step. The problem is that the trend is also a trap. Larger body size requires more energy, longer gestation, slower reproduction, and narrower habitat tolerance. When conditions change suddenly, the largest members of the lineage are the most vulnerable because they cannot downsize, cannot reproduce fast enough to adapt, and cannot survive on reduced resources. Cope's Rule (the tendency toward larger body size within a lineage is persistent and directional, not random, until an environmental shift inverts the advantage) has been confirmed across mammals, reptiles, marine invertebrates, and insects in the fossil record.
The failure mode is invisible during success. Growth solves the problems growth creates: higher revenue covers higher fixed costs, larger teams handle larger backlogs, bigger infrastructure supports bigger operations. The metrics all confirm that growing was the right decision, which encourages more growing. The trap closes when an external shock arrives that favors the opposite qualities: agility, low overhead, speed of adaptation. At that point, the accumulated size cannot be unwound faster than the shock's timeline demands. The sizing question is specific: at what point does the marginal competitive advantage of additional size become smaller than the marginal increase in extinction risk? Every lineage has an optimal size range, and the lineage that detects when it has entered the danger zone while still dominant is the one that lasts.
The decision tool: When evaluating whether growth is still serving a system or has become its primary vulnerability, ask two questions. First, how much of the current advantage depends specifically on size rather than on capabilities? If the answer is "most of it," the system is in Cope's danger zone. Second, if conditions changed tomorrow in a way that penalized size, how long would it take to reduce resource requirements by 40%? If the answer is "years," the lineage is one shock away from the fossil record.
For more than a century, neuroscience operated on a simple division of labor: neurons think, and everything else supports them. Astrocytes, the star-shaped cells that outnumber neurons in many brain regions, were classified as "neural glue," maintaining chemical balance, delivering nutrients, and cleaning up waste. Three papers published simultaneously in Science in 2025 overturned this framework with experiments across three species. In zebrafish, researchers used targeted lasers to disable astrocytes in a swimming fish. The fish never stopped. When they activated the same astrocytes artificially, the fish halted immediately. The astrocytes were not supporting the motor circuit. They were gating it, determining whether the circuit fired at all. Parallel experiments in mice and fruit flies confirmed the same architecture: astrocytes regulate calcium influx into neural circuits with reliable behavioral consequences, controlling states of alertness, anxiety, and motivation. The cells we thought were infrastructure turn out to be the regulatory layer that decides which circuits run and which stay quiet.
The finding challenges a pervasive assumption about how complex systems work: that the visible, active components are where control lives, and the infrastructure layer merely enables what the active layer decides. Neuroscience built entire brain models, including connectomes mapping every neuron and synapse, that left astrocytes out entirely. Those models predicted circuit behavior. They could not predict state, the difference between an alert brain and a drowsy one running the same circuits. The missing variable was not a neuron they had failed to map. It was the regulatory layer they had classified as plumbing.
When you are trying to change an outcome in a complex system and interventions aimed at the visible decision-makers are not producing results, stop and map one layer deeper. Identify the functions that gate what the decision layer can actually do: who controls the calendar, the approval workflow, the information flow, the resource allocation. In your own productivity, the constraint may not be your strategy but the environmental conditions, sleep, energy, context-switching frequency, that determine whether your strategic intentions translate into sustained focus. Before adding more neurons to a problem, find the astrocyte.