Two ways to build an internet. Two very different maps.
Every network is a supply chain wearing a service's clothing. Before the first byte reaches a home in Okanogan County, hardware has crossed oceans, factories, and launch pads. This page traces both machines — the satellite constellation and the county's local wireless network — from silicon to your roofline, using IPO filings, customs records, supplier disclosures, and the trade press.
Neither of these machines is American, and neither is simple. Both begin as silicon in Europe or Asia, cross at least three oceans, and end in a county of forty-odd thousand people. That much they share — and any page telling you one of them is “local” is selling something. The difference isn’t where the parts come from. It’s what each chain has to keep doing to keep you online — and who is standing next to you when it stops.
On chips, Starlink is the better-documented chain — and the better hedged. Its phased-array silicon is a decade-long partnership with STMicroelectronics: fabbed in France and Malta, packaged in Malaysia, over five million chips a day, deliberately outside the Taiwan cluster. Tarana claims custom silicon but publishes no foundry at all. Credit where it’s due.
Both chains still run through Taiwan. SpaceX has been moving its Taiwanese electronics to Vietnam and Thailand since 2024 — its own war-risk hedge, executed as logistics. Tarana’s recorded electronics shipper is Taiwanese. Same island, both machines.
One is a conveyor; the other is a stockpile. Starlink satellites live about five years, so roughly a fifth of the constellation has to be rebuilt and relaunched every year, forever — about seventy satellites a week, on rockets that flew half of everything launched from Earth last year. A tower radio has no expiry clock.
The last thousand miles is where they stop resembling each other. A Rev 4 dish has no screws — it’s glued shut, so a failure is replaced, not repaired, and the replacement arrives by parcel. The county’s radios sit on American distributor shelves by the thousand, one truck roll away.
Both answer to a distant control plane. Starlink runs on one global scheduler; Tarana’s radios require a mandatory cloud licence of about fifty-three dollars per radio per year. A fair map flags both directions — so this one does.
Where something is unknown, this page says so. Tarana’s foundry is undisclosed. Nobody has mechanically opened a Rev 4 dish. Claims that failed verification — and there were many — were cut rather than softened.
What follows is the evidence, not decoration. Every dot on the globe is a documented site and every ribbon a documented flow — spin it, isolate one layer of the chain, or break it with a stress test and watch what each machine loses.
How it works
The risk read
The flow, drawn to the real geography.
Every dot is a real place; every line is a documented flow, colored by what it carries — wide at its source, pointed at its destination. Tap any dot or line for its story. Sources at the bottom of the page.
Pick any node on the globe
Factories, fabs, launch pads, warehouses — and one valley in Washington where both chains end.
What you're seeing — exactly
Geography is real (Natural Earth coastlines; node coordinates are the documented sites). Arcs are schematic great-circle routes — they show that hardware flows between places, not the literal shipping lane. The orbital ring is drawn at a symbolic altitude.
Chain colors: gold = Starlink's chain, teal = Tarana's. Dashed service arcs are data, not freight.
Draft status: every node and arc on this globe is being verified against primary sources before publication; anything that fails verification gets cut.
The Sankey — stage by stage, and where the two chains cross.
Left to right: silicon → components → American factories and warehouses → launch or delivery → the county. Bars and ribbons are colored by region — the geography of risk — and each ribbon blends from its source region into its destination. Pick a stress-test on the globe and watch the diagram reflow: broken flows collapse, at-risk flows thin, the bars shrink to what still moves. (Ribbon widths are schematic emphasis by flow type, not measured tonnage.)
Everything at once — the flat map.
The same verified network, all visible in one frame. Note what jumps out: both chains pass through Taiwan; Starlink's chip supply deliberately avoids it (France & Malta), and its Taiwanese electronics are mid-flight to Vietnam and Thailand — SpaceX's own war-risk hedge, visible as geography.
Open them up.
The map asks where the parts come from. The hardware answers a second question: what is actually in the box, and what happens to it when it fails? Both units are drawn live in your browser from public teardown photography and published specifications. Drag to spin, pull them apart, and click any layer to read what it does — and where this page is inferring rather than observing, it says so.
What the teardown adds to the map
The dish is sealed for good. There are no screws in a Rev 4: the housing is glued about nine millimetres behind the front face, so opening one destroys it. That is a deliberate, sensible choice for a mass-produced outdoor appliance — and it means the repair path for a failed dish is a new dish, arriving the way the first one did, by parcel. The county’s radio fails differently: it is a bolted, tower-mounted unit that a crew swaps from distributor stock and carries away. Same continent-spanning supply chain behind both boxes; two completely different last hops when one of them breaks.
Now point them at the sky.
A teardown shows you the parts. It doesn’t show the thing the parts exist to do: put energy in one direction and not the others. That shape — the beam — decides who gets served, how many at once, and how much is lost simply because the customer isn’t straight overhead. Below are three antennas of genuinely different kinds, each drawn from the best public data that exists for it. Every curve is computed live from published geometry. What changes between them is how good that published input is — and each one says so.
Computed, not sketched. Each curve is the array factor summed over the real element geometry and multiplied by an element pattern — the standard calculation, run in your browser at about 2,400 angles. You can check it against theory: the dish comes out with a first sidelobe of −17.6 dB, the exact textbook value for a uniform circular aperture, and steering the array to 45° widens its beam by 1.41×, which is exactly 1 ÷ cos 45°.
Then turn it in 3D — because one flat slice quietly lies. A single cut through a beam implies the beam is round. None of these are. The Starlink array is 348 × 540 mm, so its beam is a blade, not a pencil — and when you steer it, it broadens only in the direction it is steering. The tower’s array is the extreme case: turn it and the same antenna reads as a broad fan from one side and a thin wafer from the other. That is one object, honestly drawn, looking like two different things depending on where you stand — which is exactly why the flat view alone isn’t enough.
The scale buttons are not a garnish. The same antenna, honestly plotted, changes shape enormously depending on whether the radius means power, field, or decibels — and how far down the decibel floor sits. One identical −13 dB sidelobe can occupy 4.7% or 78% of the radius, and every one of those pictures is correct. Antenna plots are routinely published without saying which was chosen. Press the buttons and watch a modest antenna become an impressive one.
Three jobs. Three different answers.
None of these is a better antenna than the others. Each is a rational answer to a completely different question, and you can read the question off the hardware.
First, clear up the thing that trips everyone. The roof dish and the tower radio are both phased arrays, and both steer electronically — neither has a moving part. Steering is not what separates them. What separates them is what the steering is for: one chases a target that never stops moving, and the other separates targets that never move at all. Only the gateway dish aims the old way, by turning the whole antenna.
A satellite crosses the sky in minutes. The dish has to re-aim continuously and hand off to the next one every few seconds, thousands of times a day, on a roof nobody will ever visit again. Anything with a motor would wear out. So: electronic steering, a single beam, and a 110° view because the target can be anywhere overhead. It is flat because flat ships in a box and bolts to a gutter — and a flat panel is the one shape you can build ten million times.
What that costs: a flat panel is worst when it looks sideways. At the edge of its view the beam is widened and about 2.6 dB weaker — SpaceX’s own filed figure, which the calculation above reproduces. Spread thinner and dimmer at once, and the footprint stretches at the same time. In a valley with ridges on both sides, sideways is where a great deal of the sky is. The design accepts that, because the alternative — a motorised dish on every roof — cannot be manufactured or maintained at that scale.
Roughly 250 fixed addresses in one 90° slice of horizon, in spectrum the operator does not own. It never has to search the sky. Its problems are the opposite ones: tell its customers apart, and survive other people’s transmitters. So it beamforms digitally — eight transmit chains, six spatial streams, several houses served at once on the same frequency — cancels interference, and deliberately works the reflections off terrain instead of fighting them.
Why it can be small: the link is kilometres, not 550. The scarce resource here isn’t gain, it’s the ability to separate people — so it spends its budget on signal processing rather than aperture. The regulator quietly agrees about direction, too: at 5 GHz the home radio is licensed for roughly ten times the tower’s radiated power (44.1 against 34.7 dBm), because the subscriber unit counts as fixed point-to-point. On this network the uplink is the favoured direction — which is precisely the direction the satellite’s welded six-to-one split starves.
A few hundred professionally staffed sites moving bulk traffic between the ground and the constellation. Motors are perfectly fine when a technician is on the payroll, and there is no scale problem to solve. So it is simply as good as an antenna gets: 2.4 m of aperture, 53.8 dBi, a beam a quarter of a degree wide, pointed by hand at one satellite at a time.
Nothing about it is electronically steered and nothing needs to be. It is also completely unusable as a consumer product — it must be aimed, mounted and maintained — which is the whole reason the thing on your roof looks nothing like it.
What each one gives up
Every one of those answers buys something and pays for it somewhere else. The last column is the one that matters here: each design’s weakness stays invisible until a particular condition arrives.
| Approach | What it buys | What it pays | When the bill arrives |
|---|---|---|---|
| Tracking arrayroof dish | Aiming with no moving parts, in microseconds, on a panel that can be built by the million and bolted to a gutter. | Gain costs silicon: roughly a thousand elements, each with its own amplifier, drawing 75–100 W continuously — near 200 W when melting snow. A reflector gets its gain from shaped metal and draws nothing. The lattice is also cut for one band, so the same spacing that behaves at 11.7 GHz is at the grating-lobe limit at 14.25. | When the satellite is low. A flat panel is weakest looking sideways — and a valley with ridges spends most of its sky there. It is also a real power bill for anyone on a generator, a well pump budget, or solar. |
| Multi-beam arraytower radio | Several houses served at once on one frequency, coverage without line of sight, and survival in spectrum nobody owns — from a small, cheap aperture. | The value moves out of the metal and into software nobody outside the company can audit, running on hardware that needs a licence to keep working — about fifty-three dollars per radio per year. It also only works because the customers never move. | When someone else changes the terms. The vendor, on the licence; or the spectrum regulator, through the databases that can order a radio to move channel or turn its power down. Neither is a hardware failure, and neither is visible in a lab report. |
| Dishgateway | The most gain per dollar anyone knows how to buy, across a huge span of frequencies, from a passive shape with almost nothing to go wrong. | Land, steel and a technician. Motors wear. A quarter-degree beam has to be pointed and kept pointed through wind and heat. And one dish is exactly one conversation. | When you need a second satellite. There is no electronic answer — you buy another dish, another mount, another patch of ground. Which is why this is infrastructure and never a consumer product. |
Read down the middle column and the three are spending different currencies for the same thing: the array spends silicon and watts, the tower spends computation and independence, the dish spends land and labour. That is not a ranking — each is the cheapest available answer to its own question. It is worth naming because the currency a company chooses is the one it then has to keep paying, and it decides what happens to you when the price of that currency moves.
How good is the evidence, per antenna?
These sit in three different evidence classes, and the difference isn’t really about secrecy — it is about which rulebook each one is licensed under. No radiation pattern has ever been published for any Starlink or Amazon customer terminal, and none ever will be: a dish bolted to a house is a licensed earth station, not certified equipment, so it never passes through the FCC test lab that would produce one. Both companies’ equipment records contain nothing but Wi-Fi routers.
Tarana G1 base node
The only genuinely measured antenna pattern in this comparison. An accredited third-party lab measured the county’s 6 GHz radio for its FCC filing — 32 polar plots, filed non-confidential, because certified equipment has to show its work.
What it still doesn’t show: those are per-element patterns. The beam Tarana’s processor actually forms toward one house appears in no public document anywhere. The envelope is public; the aiming inside it is not.
Starlink Rev 4 dish
No pattern — but SpaceX’s own patent gives the thing a pattern is computed from: a 12.3 mm triangular lattice. That, plus the panel size, is enough to calculate the beam honestly.
Where this is inferring: the array outline and element count for Rev 4 have never been published. The widely-repeated “1,280 elements” belongs to the older round dish and does not apply here.
Amazon Leo terminal
Amazon has published a customer-terminal gain range — 30 to 41 dBi — and nothing else. No pattern, no element count, no aperture. So the dish shown here is instead Amazon’s 2.4 m gateway, the one antenna it did file a real gain-versus-angle plot for.
And the terminals cannot be sold. Amazon’s own site carries the FCC notice: these devices “have not been authorized… may not be offered for sale or lease.”
The finding that isn’t about physics
Everything above is about what an antenna can do. This is about what someone decided it would do — and it is in SpaceX’s own words, in its own FCC filing.
The satellites do not use one kind of beam. The filing describes pairing “high-gain beams in highly populated areas with low-gain beams in rural areas.” Those are two distinct classes, ten decibels apart — and the arithmetic confirms it: the low-gain beam is 3.17× wider, exactly the √10 a 10 dB gain difference requires. Same spacecraft, two settings, sorted by how many people live underneath.
Say the fair version first, because it is the strong one. This is good engineering, not neglect. A wide beam covers more ground with less concentrated capacity, and that is exactly what you want where there are fewer customers per square mile. Anyone designing this system from scratch would do the same.
And that is the point. Rural service quality here is not a physical constant and not an accident of geography. It is a dial — set deliberately, from orbit, by one company, adjustable without asking anyone. Usually the “distant control plane” on this page is a software story: a global scheduler, a mandatory cloud licence. The beam classes make it physical. Not the price, not the terms — the watts arriving at a roof.
Count the units and the factory follows
Each of those three answers implies a completely different thing to build, and that is where this page’s two maps come from. Ten million roof panels at a thousand-odd elements each is a semiconductor commitment — which is why Starlink’s chain runs through a decade-long STMicroelectronics partnership turning out five million chips a day in France and Malta. The antenna is the demand. A few thousand base nodes at eight chains each is a software commitment: the value sits in beamforming code, which is exactly the part Tarana will not publish and no lab report can show you. A few hundred gateway dishes you can simply buy from a dish maker. One firm’s advantage sits in a fab it does not own; the other’s sits in software it will not describe; the third needs neither. That is visible in the shape of the beam before you ever reach the parts list.
The stress test — named parts, real dates, honest unknowns.
Maps show structure. This section adds judgment — using the methods the field actually trusts: MIT’s time-to-recover stress test, the White House semiconductor review’s concentration analysis, and the three-part chokepoint test used in national-security supply-chain work. Every number below is documented; everything undisclosed is scored undisclosed, never guessed.
The ledger asks two questions per node, from the stress test MIT’s David Simchi-Levi built with Ford (Harvard Business Review, 2014): how long would this link take to rebuild if it vanished (time-to-recover), and how long would the county stay served without it (time-to-survive)? No war probabilities are invented — if rebuild outruns survival, the exposure is real at any odds. Ford’s famous finding applies here too: the dangerous parts are rarely the expensive ones.
A “chokepoint” must pass three tests (the CSIS criteria): one supplier dominates, no alternative exists today, and the dependence continues after the sale — spares, software, support. Note what vertical integration does under this test: it doesn’t erase a chokepoint, it changes its owner.
The chart placements are our judgment from the documented record — and the reasoning for every dot is printed beside it, because a score without its rubric is just an opinion with an axis.
Rebuild time vs. survival time
The whole comparison in one table: if this link snapped tomorrow, how long to rebuild it — and how long before the county feels it. Bars are schematic (days → years, left to right); the words carry the evidence.
| Link | Time to recover — rebuild it | Time to survive — without it | Reading |
|---|---|---|---|
| The launch conveyorStarlink | Documented stand-downs: 177 days (2015), 135 days (2016) — then 3 to 15 days across three FAA groundings in a single summer of 2024. Mature-era recovery is fast; it is also frequent. 3 days – 6 months, documented range | The constellation coasts — and thins. A ~5-year satellite life means ~20% of the fleet needs replacing every year, so a long halt shows up as shrinking capacity, not a blank screen. years — degrading the whole time | Clock running Degrades, doesn’t stop — and never stops needing launches. |
| ST phased-array chipsStarlink | No second source disclosed. The White House semiconductor review documents why swaps are slow: only qualified producers, “months to switch a production line,” manufacturing alone “up to 26 weeks” — and a modern fab costs $12–20 billion. SpaceX’s disclosed answer is Terafab, its chip venture with Intel. years-class — fabs don’t improvise | Buffer undisclosed — and precedent says thin: in November 2021 SpaceX told its own customers “silicon shortages have delayed production,” and Starlink signed only 5,000 subscribers in the last two months of that year. Installed dishes keep working; the shock lands on new kits and replacements. undisclosed — 2021 precedent was months of delays | Undisclosed Named partner, unnamed buffer. |
| Bastrop terminal plantStarlink | One company-owned site builds all current-gen kits (Hawthorne runs a second terminal line). Rebuilding a tens-of-thousands-a-day line is a months-class problem; no figure disclosed. months-class, single primary site | There is no spares layer between the factory and your roof — fulfillment is a parcel. A county household with a broken dish waits exactly as long as the conveyor takes. days of pipeline, no local stock | Exposed No local buffer — by design. |
| Accton electronicsTarana | Contract manufacturing can move — Taiwan’s ODM bench is deep (WNC, Sercomm, Gemtek…) — but a transfer is a multi-month, dozens-of-steps qualification. No published norm exists, and we say so. multi-month, alternatives exist | Documented: 3,345 remote-node bundles and ~170 base nodes on US distributor shelves the day we checked — years of spares at county scale. Installed radios have no expiry clock. years, at county scale — counted, not claimed | Buffered The shelf absorbs the shock. |
| Tarana Cloud SuiteTarana | There is no alternative cloud. The license is mandatory, and Tarana’s own terms say the hardware requires it “to operate properly.” no substitute exists | If the vendor’s cloud went dark, the county’s radios — local as they are — would be in unsupported territory. short — and that’s the point | Exposed The Tarana chain’s sharpest edge — printed in bold on purpose. |
| Tarana, the companyTarana | Replacing a dead vendor means overlaying different radios (Cambium, Ubiquiti, RADWIN…) — a documented downgrade in tree-and-terrain performance, not an impossibility. Months to years at network scale. months–years, downgrade not death | Precedent is kinder than feared: Airspan ran straight through Chapter 11 in 2024 “without disruption” and emerged in six months; Mimosa’s operators kept support through two changes of owner. Installed gear outlives its vendors. years — industry precedent, cited below | Buffered Gear outlives vendors — the record says so. |
Four chokepoints, dossier’d
The field’s standard for claiming a chokepoint (ASML, Ukrainian neon) is a five-slot dossier: name the facility, state the share, explain why the concentration exists, cite the shock that already happened, and price the alternative. A “risk” that can’t fill the slots doesn’t get published. Two per chain:
STMicroelectronics — the sole chip partner
Starlink chain · chips- The facility
- ST’s fabs in France and Malta; packaging and test in Malaysia.
- The share
- Sole supplier of the phased-array chips in every dish and every satellite — 5M+ a day, 7.5 billion shipped.
- Why it exists
- A decade of co-design. Silicon this custom has no off-the-shelf substitute — the partnership is the product.
- The precedent
- November 2021, SpaceX to its own pre-order customers: “Silicon shortages have delayed production which has impacted our ability to fulfill orders.”
- Time to alternative
- Terafab — the disclosed Intel venture — is the alternative under construction. The White House review prices this fab class at $12–20B and documents months just to switch a production line.
Launch — the chokepoint SpaceX owns
Starlink chain · build & launch- The facility
- Three Falcon pads (two in Florida, one at Vandenberg) plus Starbase for Starship — 165 launches in 2025, half of everything launched from Earth.
- The share
- 100% of Starlink mass flies on SpaceX’s own rockets — and V3 satellites are planned for Starship, which SpaceX says carries ~20× the constellation capacity of a Falcon 9 per launch.
- Why it exists
- Vertical integration. It made launch cheap — and it converts an external chokepoint into an owned single point of failure, self-insured and self-inspected.
- The precedent
- Stand-downs of 177 days (2015) and 135 days (2016); in 2024, three FAA groundings in three months — 15 days, 3 days, and ~2 weeks.
- Time to alternative
- None at this cadence — no other launcher on Earth could absorb the conveyor. By Aug 2026 trade press described satellite operators “in panic” over launch scarcity — a shortage that binds everyone except the company that owns the rocket. And the dependence is narrowing to one vehicle: SpaceX is winding the Falcon 9 down (Shotwell: operational life may end ~2030–32; Ars reports, via one source, commercial sales ending after 2028), so a long Starship delay parks not just the V3 growth story on the pad with it, but eventually the whole conveyor.
The undisclosed silicon — and Accton
Tarana chain · chips & electronics- The facility
- Design in Milpitas; US customs records show Accton Technology (Taiwan) as the recurring electronics shipper. The foundry behind the “custom silicon” is unnamed.
- The share
- One recorded electronics source, and claimed custom silicon in every radio — with no published fab and no verified teardown.
- Why it exists
- Standard fabless-startup economics: one qualified ODM, one silicon spin at a time. Normal — and opaque.
- The precedent
- 2021–22: market-wide chip lead times peaked at ~27 weeks (Susquehanna data), with RF front-end parts on allocation across the industry.
- Time to alternative
- Taiwan’s ODM bench is deep (WNC, Sercomm, Gemtek…), but a transfer is a multi-month qualification — and losing the foundry would gate everything behind a silicon re-spin. The least documented layer on this page, scored accordingly.
Tarana itself — protocol, cloud, viability
Tarana chain · the vendor- The facility
- A private Milpitas company, ~$426M raised — no public financials, and no distress signals found either. Both facts belong in the ledger.
- The share
- 100% of the protocol. The radios interoperate with nothing else, and the mandatory cloud license (~$53 per radio per year) continues the dependence after the sale — the textbook “leverage” criterion.
- Why it exists
- Proprietary performance is the moat and the lock-in — one fact, seen from two sides.
- The precedent
- The gear-outlives-vendor record: Airspan operated straight through its 2024 Chapter 11 “without disruption”; Mimosa’s operators kept support through two owners; Cambium — the most-named alternative — spent 2025–26 in restatement and delisting proceedings while its networks kept running. Vendor distress is normal here; dead networks are rare.
- Time to alternative
- A full overlay with Cambium / Ubiquiti / RADWIN-class gear: months to years, and a documented downgrade in non-line-of-sight coverage. Replaceable — at a price.
Every part on one chart
The purchasing classic (Kraljic, 1983): plot each dependency by how badly its loss hurts the service against how hard it is to substitute. The top-right corner is where chokepoints live — and the bottom-left is printed too, because an analysis that only shows danger isn’t calibrated.
- ST phased-array chips — sole partner, in every dish and satellite; no second source disclosed; fab-class alternatives cost $12–20B and years.
- The launch conveyor — 100% self-launched; no other provider on Earth could absorb the cadence; stand-downs are documented and recurring.
- Redmond satellite line — one factory feeds the entire replacement treadmill; a long outage echoes in orbit for five years.
- Bastrop terminal plant — company-owned single primary site, but Hawthorne runs a second terminal line — that fallback moves it left of the chips.
- Filtronic gateway amps — one named UK firm, but gateways are a buffered, 400-site layer; degradation would be gradual.
- Routers (WNC, Vietnam) — carrier-grade router assembly has a deep global bench; annoying to move, not hard.
- Undisclosed custom silicon — in every radio, foundry unnamed; a foundry loss gates everything behind a re-spin. Sits in the corner because it’s undocumented.
- Tarana itself — proprietary protocol plus mandatory cloud is the textbook leverage chokepoint; the mitigations are the stocked shelves and the gear-outlives-vendor record.
- Accton electronics — single recorded source today, but Taiwan’s ODM bench offers real alternatives after a multi-month transfer.
- Distributor shelves — two national warehouses plus a regional affiliate, thousands of units deep: the anti-chokepoint.
- Mounts & accessories — brackets are not phased arrays. Printed here so you can see the chart is calibrated, not alarmist.
Concentration, stage by stage
The White House review’s core move, applied to both chains: walk the stages and classify each one. Sole source means no alternative exists; single source means one was chosen but alternatives exist — opposite remediation stories, so we never blur them.
| Stage | Starlink (SpaceX) | Tarana (county network) |
|---|---|---|
| Silicon | Sole source STMicroelectronics — disclosed, decade-long, deliberately outside Taiwan; Terafab venture is the hedge under construction. | Undisclosed “Custom silicon,” foundry unnamed — unscoreable, which is itself a score. |
| Radio / terminal build | In-house, single primary Bastrop builds all current-gen kits (Hawthorne second line); satellites from one Redmond factory. | Single source Accton (recorded) — with a deep Taiwanese ODM bench behind it (WNC, Sercomm, Gemtek…). |
| The last thousand miles | In-house parcel Factory → courier → your roof. No local stock, no local hands. | Multi-source Two national distributors plus a Washington-state affiliate, with counted stock on shelves. |
| Operations & control | One operator One global scheduler from your roof to the backbone; terms changeable unilaterally. | Vendor cloud Mandatory license (~$53/radio/yr) on a private vendor’s cloud — the chain’s sharpest edge, flagged in the ledger above. |
Read the last row again — it’s the honest symmetry of this page: both systems answer to a distant control plane. The asymmetry is everywhere else — one chain ends at a parcel and a help line, the other at a stocked shelf and a bucket truck.
Who actually decides.
A supply chain is a set of decisions before it is a set of shipments. When a link snaps, somebody chooses what gets built, what gets funded, what gets shipped first — and which customers wait. This section maps that authority, using the same standard as the rest of the page: public record only.
This is key-person analysis, not character assessment. Concentration of decision-making is a standard supply-chain risk: it is the human version of the single-source question asked of chips and factories above. Where a chain’s critical judgment, capital, or architectural knowledge sits with very few people, that is an exposure worth naming — and it is often named by the companies themselves.
Sourcing rules. Roles and histories come from company filings, official biographies, press releases, and credible trade reporting. Nothing here touches anyone’s private life, family, health, or personal politics. Anonymous employer-review and forum material was not used: it cannot establish anything about a named individual. Where a title is disputed between sources, the dispute is printed rather than resolved.
Two of the portraits here are photographs; three are illustrations, and each is labelled as one. Musk and Shotwell are public-domain US federal photographs. Alwan and Hill are illustrated from published photographs, used with those two executives’ permission. Johnsen is an illustrated likeness drawn from published photographs — a drawing, not a picture of a moment that happened. Everyone else carries initials, because we do not republish headshots that are licensed to their owners rather than to us.
One asymmetry is itself a finding. Tarana publishes a full executive roster with biographies. SpaceX — a company about to be worth more than a trillion dollars — names very few operating leaders publicly; below its officers, the public record thins fast. The reader deserves to know that the two columns below are not equally documented, and why.
The county’s vendor
Tarana is a private company of roughly five hundred people whose leadership is, unusually, a reassembled version of two prior teams: the TiMetra/Nokia carrier-router group, and the BRN Phoenix/BeamReach fixed-wireless group. That continuity is an argument for execution capability — and a concentration of it.
Basil Alwan
Chairman and Chief Executive Officer, since 2021
Tarana- What he controls
- Strategy, financing, pricing, and how long the company supports what it has already sold. He holds both the chair and the chief executive seat, so there is no independent board chair sitting across from the CEO.
- Track record
- Founded TiMetra Networks in 2000; Alcatel acquired it in 2003 (price never disclosed). That carried him into leading Nokia’s IP/Optical Networks business. Tarana credits him with building the first IP/MPLS service router to pass $3 billion in annual revenue — a company-stated figure we could not independently verify.
- The key-person read
- Carrier-infrastructure experience at real scale is exactly what a county wants behind gear it will run for a decade. The exposure is the reverse of the same coin: combined chair/CEO authority in a venture-backed private company, with no public financials to check it against.
Andy Hill
Vice President of Operations
Tarana- What he controls
- By Tarana’s own description, “all aspects of manufacturing and distribution” — which makes him the single named person accountable for everything the map above draws on the Tarana side: the contract manufacturer, the shipping lanes, and the stock on American distributor shelves.
- Track record
- Twenty-plus years in supply chain, manufacturing, test, and delivery logistics; previously COO of PlantSense, with operations roles at Nokia, BRN Phoenix, Diamond Lane, and Stanford Telecom.
- The key-person read
- This is the most supply-chain-relevant name in either company, and the most reassuring one on this page: the function is named, senior, and staffed by someone whose whole career is manufacturing logistics. What remains unknown is the thing no bio can answer — whether a second contract manufacturer is qualified and waiting.
Dale Branlund
Co-founder. Listed by Tarana simply as “Co-Founder”; his biography describes him as chief architect of earlier generations of wireless systems.
Tarana- What he controls
- The physical-layer architecture that is Tarana’s entire commercial claim — the beamforming and interference-cancelling behaviour that lets a radio work without line of sight.
- Track record
- Forty years in adaptive arrays and signal processing, at BRN Phoenix, BeamReach Networks, and Radix Technologies. Co-inventor of the first wireless system to combine OFDMA with adaptive beamforming — AT&T’s Project Angel. The credential is real and it is the direct ancestor of what the county bought.
- The key-person read
- This is the sharpest key-person exposure in the Tarana chain, and it compounds the undisclosed-foundry problem in the stress test above: the moat and the risk are the same fact. Deep architectural knowledge concentrated in a small founding group is why no competitor has copied the product — and why a small number of departures would matter more here than at a large vendor.
Sergiu Nedevschi
Co-founder. Title disputed: Tarana’s own site lists him as CTO; his LinkedIn profile currently reads Chief Strategy Officer. We print the conflict rather than pick.
Tarana- What he controls
- Product and technology direction alongside Branlund, from the company’s founding.
- Track record
- Wireless research at Intel Labs and the International Computer Science Institute; designed and deployed WiLDNet with UC Berkeley’s TIER group, which built long-range wireless for places the market had skipped. Twenty-five-plus published papers.
- The key-person read
- A research pedigree aimed squarely at hard rural links is a good sign for a county network. Note the honest gap: we found no public statement from him about the platform’s architecture, so nothing is attributed to him here beyond his documented record.
Carl Guardino
Vice President of Government Affairs & Policy, since November 2022
Tarana- What he controls
- Tarana’s influence over the funding rules that decide whether fixed wireless is eligible for public broadband money — the same BEAD rules this county lives under.
- Track record
- Twenty-four years as chief executive of the Silicon Valley Leadership Group, then global government affairs at Bloom Energy. He has led or co-led nineteen ballot initiatives and won eighteen. He also currently chairs the California Transportation Commission, a state body allocating billions annually — a public role held at the same time as a vendor’s lobbying role. Both facts are public record; we found no allegation of impropriety in either.
- The stated agenda
- Openly declared, not hidden: technology-neutral BEAD funding, against what Tarana argues would be a $200 billion-plus fiber-only approach. In his words, high fiber costs and long timelines “will leave millions of unserved.” Read every Tarana claim about BEAD, fiber, and satellites knowing a paid advocate is making it — including the claims this page finds persuasive.
Sri Reddy
President and Chief Operating Officer
Tarana- What he controls
- Day-to-day execution beneath the chief executive: engineering delivery, support, and the service-provider relationships that decide whether a rural ISP can actually get a failed radio replaced.
- Track record
- Twenty-plus years at Nokia, most recently running its $5 billion IP/Optics division, and before that a founder and engineering leader at TiMetra alongside Alwan. One clarification the page owes you: that $5 billion business is the same Nokia unit associated with Alwan — it is one business, not two.
- The key-person read
- The reassembly of a proven carrier team is genuine execution evidence. It also means the top of this company is a small group of people who have now worked together for two decades.
The satellite operator
SpaceX discloses less about who runs Starlink than Tarana does about who runs Tarana — but on the question that matters most here, it does the work for us: the company’s own stock-offering filing states the dependency in writing.
Elon Musk
Founder, Chief Executive Officer, Chief Technical Officer, and Chairman of the Board
SpaceX- What he controls
- By the company’s own filing: approximately 82.4% of the voting power. Class B shares carry ten votes each; Class B holders elect a majority of the board, and he alone “will be able to elect, remove or fill any vacancy” among them. SpaceX qualifies as a “controlled company” under Nasdaq rules and may therefore opt out of requiring a majority-independent board. One person holds the votes, the board majority, the chief executive seat, and the chief technical seat.
- What the company says about that
- These are SpaceX’s words, not ours. It is “highly dependent on the continued services of Mr. Musk”; “We do not maintain key-person life insurance on Mr. Musk”; and although he is “highly active in our management,” “he does not devote his full time and attention to our businesses” — the filing lists Tesla, Neuralink, The Boring Company, and a previous role as Senior Advisor to the President. The charter also states he and his affiliates are “not restricted from owning assets or engaging in businesses that compete directly or indirectly with us.”
- On the record
- In September 2022 Ukraine asked SpaceX to extend Starlink coverage to Sevastopol for a naval-drone operation, and he declined. Walter Isaacson’s biography first described this as secretly ordering coverage switched off; Isaacson corrected that account on 8 September 2023, and the Washington Post appended a correction: coverage within 100 km of the Crimean coast had already been geofenced off, and he refused a request to activate it. The corrected version is the one that matters here — and it is still a single person deciding where a network works during a war.
- The key-person read
- The strongest key-person disclosure on this entire page was written by SpaceX’s own securities lawyers. Apply the stress test above to it: the recovery time for this node is undisclosed, and the company states plainly that it carries no insurance against it.
Gwynne Shotwell
President and Chief Operating Officer since 2008; director since 2009
SpaceX- What she controls
- Day-to-day operations and the customer and government relationships that turn strategy into flown missions. Worth noting precisely: the filing documents one concrete decision right — she and Musk historically set the chief financial officer’s salary. Beyond that, her authority is real but not separately enumerated in the offering document.
- On the record
- At the FAA’s commercial space conference on 8 February 2023, she said SpaceX had restricted Ukraine’s ability to use Starlink for drone control, saying the service was never intended for offensive use and that Ukrainians had leveraged it “in ways that were unintentional and not part of any agreement.” Asked whether SpaceX had acted, she said: “There are things that we can do and have done” — and declined to specify what.
- The key-person read
- She is the institutional counterweight, and eighteen years in the seat is genuine operational continuity. But the counterweight is also one person: a company defended against key-person risk by a single deputy has two key people, not none.
Bret Johnsen
Chief Financial Officer since 2011
SpaceX- What he controls
- Where the money goes — long-term financial strategy and the financial side of every growth initiative, now across launch, the constellation, and an AI business inside the same company.
- The fact worth reading twice
- The filing says SpaceX expects roughly $74.4 billion of net proceeds, to be used for “the expansion of our AI compute infrastructure, enhancements to our launch infrastructure and launch vehicles, increases in the scale and capacity of our satellite constellations.” Read the order: AI compute is named first; the satellite constellation third. That is not a prediction or an inference — it is the order the company chose in its own offering document, and management retains “significant flexibility in applying the net proceeds.”
- The key-person read
- For a county whose service depends on that constellation being continuously rebuilt, capital priority is the quiet risk. Nothing here suggests the constellation is being starved. It does mean the conveyor now competes for money with a data-centre buildout under the same roof.
Michael Nicolls
Title varies by source: the U.S. Commerce Department’s space advisory committee lists him as Vice President for Starlink Engineering; trade press since early 2026 calls him Starlink SVP. No promotion was ever announced, so we print both.
SpaceX- What he controls
- Per that federal biography: the teams developing and operating Starlink broadband and Direct-to-Cell — satellites, user terminals, ground technology, the network software, and constellation operations. That is essentially everything this page has drawn: the Redmond output, the dish in the teardown, the 400-plus gateways, and the scheduler.
- Track record
- Founder and chief technical officer of LeoLabs, which tracks objects in orbit and models collision risk; Cornell doctorate in electrical engineering. A space-safety specialist holding the constellation’s engineering is a genuinely reassuring fact.
- On the record
- On 1 January 2026 he announced SpaceX would lower roughly 4,400 satellites from about 550 km to about 480 km across 2026 — lower orbits decay far faster, so dead hardware clears itself in months rather than years. A decision that trades some operating margin for less debris is exactly the kind of judgment worth crediting.
- The key-person read
- Enormous technical scope on one person, and Business Insider reported in April 2026 — from an internal memo, with no company confirmation and no filing — that he had also become president of xAI while remaining at Starlink. We flag it as single-source. If accurate, the divided-attention pattern the filing discloses about the chief executive would repeat one level down, at the person holding the constellation.
The people who run the factories
John Federspiel · Akash Badshah · Cornelia Rosu — senior directors, named nowhere in the stock filing
SpaceX- What they control
- John Federspiel, senior director of Starlink product engineering, runs the Bastrop terminal factory — on camera in SpaceX’s own video: “Raw plastic pellets come in, raw aluminum comes in… right now, we’re producing 15,000 a day.” Akash Badshah, senior director of satellite engineering, speaks for the Redmond satellite factory — the ~70-a-week line and the laser links. Cornelia Rosu is senior director for Starlink production.
- Why they share a card
- Because the public record will not support individual profiles, and we would rather say so than pad three cards with LinkedIn scrapes. Their titles are known largely because SpaceX put them in its own factory videos.
- The disclosure read
- Here is the finding: SpaceX’s offering document — a filing of more than a million and a half characters — names exactly three executive officers: Musk, Shotwell, and Johnsen. The people running the two factories the entire county-side argument depends on do not appear in it at all. Tarana, a private company a fraction of the size, publishes eighteen executive biographies. On knowing who makes the hardware, the smaller company is the transparent one.
A 2023 Reuters investigation documented at least 600 previously unreported worker injuries at SpaceX since 2014, including eight amputations and one death — Lonnie LeBlanc, killed at McGregor, Texas in 2014, after which OSHA found the company failed to protect from a clear hazard and fined it $7,000. Reuters calculated 2022 injury rates per hundred workers of 4.8 at Brownsville and 2.7 at McGregor against a space-industry average of 0.8.
What that record does not establish, and we will not imply: the industry-average comparison is Reuters’ own analysis, not an agency finding; the 600 figure is a journalistic compilation, not an official tally; the rates are not normalised for the kind of work done at each site; and no regulator has found systemic illegality. It is not evidence of defective hardware. We include it because manufacturing culture is a legitimate supply-chain question when a company’s whole model is building at speed — and we bound it because the page’s standard is what is documented, not what is implied.
What the map actually says
Both chains are global — that's electronics, and pretending otherwise would be propaganda. Tarana's radios are designed in Silicon Valley and built by Asian electronics manufacturers; its mounts and power bricks come mostly from China and Vietnam. Starlink's chips are European, its routers Vietnamese, its gateway amplifiers British. Neither company could build its product from a single country. The honest question isn't "which chain is domestic?" — neither is. It's what each chain demands of the world per year of service delivered, and what happens to your internet when a link snaps.
The chip layer — where the deepest risk lives
Starlink's phased-array chips are a confirmed, decade-long partnership with STMicroelectronics: co-designed in France and Italy, fabricated in ST's own fabs in France and Malta, packaged and tested in Malaysia — running at over five million chips a day, more than 7.5 billion shipped, in every dish and every satellite including V3. That design choice quietly says something: the most chip-hungry consumer-hardware program on Earth sourced its silicon outside the Taiwan fab cluster. SpaceX's S-1 still warns its "direct chip suppliers are dependent on a concentrated group of advanced semiconductor fabrication facilities" — and discloses a chip venture with Intel ("Terafab") to pull even fabrication in-house. Tarana's silicon story is honestly murkier: the company builds its moat on "custom silicon" but publishes no foundry, and no teardown we could verify names one. What's documented is the electronics assembly layer (Taiwan's Accton in the customs records) — so we treat Tarana's chip layer as undisclosed, presumably Asian-fabbed, and score it accordingly. On chips, Starlink's chain is better documented and better hedged. Credit where due.
The crossover — Taiwan is in both chains
The single most important shared fact on the map: both machines route through Taiwan. Starlink's router/electronics tier was Taiwanese (Wistron NeWeb, Universal Microwave, Chin-Poon) until SpaceX — "mostly due to geopolitical considerations," per Reuters — asked those suppliers to move to Vietnam and Thailand, and Nikkei reports it is now purging Chinese parts and personnel chain-wide. Tarana's recorded electronics shipper is Taiwan's Accton. The difference under a Taiwan shock is not who hurts — both do — it's what hurts: Starlink's exposure is its consumables conveyor (new kits, new satellites, forever), while Tarana's exposure is new-build hardware only. An installed Tarana network keeps running with the spares already on US shelves; an installed Starlink constellation keeps consuming itself at ~20% a year no matter what.
The structural difference — a conveyor versus a stockpile
Starlink's supply chain never gets to rest: the S-1 documents ~70 satellites a week from Redmond, ~200,000 terminals a week company-wide, 400+ gateway stations, and a five-year satellite life — meaning roughly a fifth of the constellation must be rebuilt and relaunched every year, forever, through Falcon and eventually Starship. It is a magnificent conveyor — and the service exists only while the conveyor runs. Tarana's chain is a conventional telecom stockpile: build hardware, ship it to distributors (3,345 remote-node bundles and ~170 base nodes sat on one distributor's US shelves the day we checked), bolt it to towers, and let it run for a decade with spares a freight-truck away. Neither model is "better" in the abstract — but they fail differently, and for a county betting its only connection, how a system fails is the whole question.
Same table, both chains
| Dimension | Starlink (SpaceX) | Tarana (county network) |
|---|---|---|
| Design | Hawthorne / Redmond / Bastrop, TX+CA+WA | Milpitas, CA + Pune, India |
| Chips | STMicro — fabs in France & Malta, packaging Malaysia; 5M+/day; Intel "Terafab" venture disclosed in S-1 | "Custom silicon" (Tarana's claim); foundry undisclosed |
| Hardware build | Terminals: SpaceX's own Bastrop plant (all current-gen kits, tens of thousands/day) + Hawthorne; routers: Wistron NeWeb (Vietnam). Satellites: Redmond, ~70/week | Asian contract manufacturing (Taiwan's Accton in customs records); accessories China/Vietnam/Philippines |
| Named concentrated dependencies | ST (chips) · Filtronic UK (gateway amps; SpaceX holds warrants up to 15% of the firm) · launch ranges | Undisclosed radio ODM · distributor channel (multiple: WAV, Digicomm, ISP Supplies) |
| What the county buys | $599-class consumer kit, shipped parcel; no local spares, no local repair | PUD buys BNs (~$19–24.5k list) + RNs (~$170–290 each in 5-packs) via US distributors; thousands in domestic stock |
| Steady-state demand | Rebuild ~20%/yr of constellation + terminal churn + launch cadence — a permanent conveyor | Spares + growth only; installed base has no expiry clock |
| Local recovery | Reorder a kit; wait for the parcel | Truck roll from local stock; RMA behind it |
| Software tether | Starlink's cloud, one global scheduler | Tarana Cloud Suite — mandatory SMS license (~$53/RN/yr); a real dependency, honestly noted |
What this means for the county
Run the scenarios above and one pattern repeats. Shocks that hit manufacturing — Taiwan, China, a pandemic — slow new hardware for both chains. But the county's local network degrades like a stockpile: gradually, visibly, with spares in Illinois and Colorado and a bucket truck in Okanogan. The satellite service degrades like a conveyor: invisibly at first, then all at once, because the constellation is always five years from disappearing without resupply. And when the shock is local — the fire, the windstorm — the difference inverts into the county's favor entirely: supply chains with a human being at the last hop are the ones that put a valley back online. Neither chain is domestic. Only one of them ends with a neighbor.
Sources
Primary. SpaceX Form S-1/A (SEC, June 3 2026 — facilities, Bastrop/Redmond/Hawthorne/McGregor quotes, "200,000 terminals per week," "over 400" ground stations, supplier-risk language, Terafab); STMicroelectronics × SpaceX joint release (Dec 19 2025; ST 6-K) and Advanced Television (May 7 2026 — 7.5B chips); Filtronic plc announcements (Apr 24 2024 strategic partnership; Aug 26 2025 record £47.3M order; warrant disclosures); Tarana Wireless — About, G1 datasheet rev 2024-10-17, G2 launch release (Sep 4 2025), Standard T&Cs (Jul 12 2025), funding releases (2019–2023); WAV public catalog (prices and stock, retrieved Aug 2 2026); ImportYeti customs summaries (Tarana Wireless, 132 sea shipments through Jul 9 2026). Secondary. Reuters (Nov 5–6 2024 — Taiwanese suppliers asked to move; Sep 27 2024 — $1.5B Vietnam plan); Nikkei Asia (Jul 30 2026 — Chinese nationals/parts excluded); GeekWire (May 21 2026 — S-1 satellite rate); Teslarati/Space.com (Bastrop rates); Business Wire (Tarana × WAV 2022; funding 2019); Digicomm release (Feb 29 2024); Linde / City of Brownsville (air-separation plant). Stress-test section. Method: D. Simchi-Levi, W. Schmidt & Y. Wei, "From Superstorms to Factory Fires" (Harvard Business Review, Jan–Feb 2014 — the time-to-recover / time-to-survive test, built with Ford); P. Kraljic, "Purchasing Must Become Supply Management" (HBR, 1983 — the impact × substitutability quadrant); G. Allen, "Choking Off China's Access to the Future of AI" (CSIS, Oct 2022 — the dominance / irreplaceability / leverage chokepoint criteria); White House, "Building Resilient Supply Chains" 100-Day Review (June 2021 — fab costs $12–20B, "months to switch a production line," chips "up to 26 weeks" to manufacture, Taiwan 92% of leading-edge production). Launch stand-downs: CRS-7 (Jun 28 2015 → return to flight Dec 22 2015, 177 days) and Amos-6 (Sep 1 2016 → Jan 14 2017, 135 days), per launch records; 2024 FAA groundings (Jul 12–27; Aug 28–31; Sep 28–Oct 11) per FAA statements and trade press; 2025 launch total (165 of 330 worldwide) per public launch logs (J. McDowell) and Wikipedia's launch lists; BryceTech's tally counts 324 worldwide — catalogs differ on failed and marginal launches. Launch-market crunch, Falcon 9 wind-down (Shotwell's ~2030–32 on the record; the post-2028 commercial cutoff attributed to one unnamed source), and Amazon's throttled satellite production: E. Berger, "Satellite operators are in panic mode due to a worsening launch crisis," Ars Technica, Aug 2026. Chip-shortage precedent: SpaceX customer notice, Nov 2021 ("Silicon shortages have delayed production…"), via Ars Technica/Business Insider; peak lead times ~27 weeks (May 2022), Susquehanna Financial data via press. Vendor-failure record: Airspan Ch. 11 (filed Mar 31 2024, emerged Oct 2024, Delaware docket); Mimosa sold to Radisys/Jio ($60M, Mar 2023); Cambium restatement and Nasdaq delisting proceedings (2025–26, SEC filings); Siklu acquired by Ceragon (2023). ODM bench: WNC, Sercomm, Gemtek, Senao company disclosures. Quadrant placements and bar lengths are editorial judgments from these documents; each placement's reasoning is printed beside it. Profiles section. Roles and biographies: Tarana Wireless leadership page (taranawireless.com/about); Business Wire, Nov 15 2022 (Guardino appointment); Light Reading and Fierce Network (Guardino BEAD quotes); SpaceX Form S-1/A (SEC) for officer titles, voting control, and the company's own key-person and conflict-of-interest risk factors. Portraits are public-domain US federal photographs only: Elon Musk — U.S. Air Force photo by Trevor Cokley (U.S. Air Force Academy, 2022), public domain via Wikimedia Commons; Gwynne Shotwell — NASA photo by Bill Ingalls (2018), public domain via Wikimedia Commons. Both were cropped and converted to greyscale. The remaining three portraits are illustrations, labelled as such on each card: Basil Alwan and Andy Hill are illustrated from published photographs and used with those executives' permission; Bret Johnsen is an illustrated likeness drawn from published photographs rather than a photograph of any particular moment. No headshot licensed to a company or news outlet was reproduced — where we hold neither rights nor permission, initials are shown instead. Anonymous employer-review and forum material was not used for any statement about a named individual. Method. Every claim on this page was re-verified against the sources above in Aug 2026; claims from earlier research that failed verification (alleged Chinese SpaceX suppliers, a Tarana–CommScope acquisition, "Eutelsat owns 3% of SpaceX," Honeywell/Carpenter parts, NVIDIA GPUs on satellites) were excluded, and several appear on no reputable record at all. Where something is unknown — Tarana's foundry, SpaceX's fulfillment centers — this page says "undisclosed" rather than guessing.