THE THROTTLE
How Two Israeli Companies, Three Choke Points, and One Closed Strait Just Hijacked the Entire AI Build-Out
And why the consumer electronics market is being quietly killed to keep Nvidia fed.
I. Two Buildings You’ve Never Heard Of
Every chip in every AI server in every hyperscaler data centre on Earth passes through two buildings.
One sits in Migdal HaEmek, in Israel’s industrial north, an hour’s drive from the Lebanese border. The other sits in Rehovot, twenty kilometres from Tel Aviv, twelve minutes’ flight time from any Iranian missile that gets through.
Together, those two buildings produce the inspection and metrology tools without which Samsung and SK Hynix cannot ship a single stack of High Bandwidth Memory. Not “cannot ship efficiently.” Cannot ship. The yields collapse. The TSV alignment fails. The wafers go in the scrap bin.
Camtek. Nova. Two companies you’ve never heard of, sitting in a country at war, controlling the throttle on the AI boom.
Wall Street is watching the wrong choke point. The story everyone is telling about the 2026 semiconductor crisis is helium from Qatar and naphtha from the Strait of Hormuz. That story is true. It is also incomplete. Beneath it sits a deeper, more concentrated, more permanent vulnerability that no major analyst has connected to the current crisis: the entire AI memory supply chain runs through Israeli measurement instruments that cannot be substituted on any timeline that matters, made by companies whose engineers are currently being mobilised for reserve duty on three fronts.
This is the article about the throttle.
II. The Peninsula That Makes the World
Begin with the scoreboard. South Korea, a peninsula the size of Iceland, produces approximately seventy per cent of the world’s DRAM, fifty-three per cent of the world’s NAND, and somewhere between eighty-five and ninety per cent of the world’s High Bandwidth Memory. Two companies, Samsung Electronics and SK Hynix, account for almost all of it. They are not the supply chain. They are the supply chain.
In late October 2025, in a private signing ceremony in Seoul witnessed by President Lee Jae-myung, OpenAI’s Sam Altman, Samsung’s Jay Y. Lee, and SK Group’s Chey Tae-won put their names to Letters of Intent committing the two Korean memory makers to provide nine hundred thousand DRAM wafer starts per month for the Stargate AI infrastructure project. Nine hundred thousand. That is approximately forty per cent of total global DRAM wafer capacity, dedicated to a single customer, for a single application, on contracts whose total value through 2029 has been estimated by Korean analysts at over one hundred trillion won, or seventy-two billion US dollars.
Two months later, in early December, SK Hynix’s Q4 2025 earnings revealed the company had effectively sold out its entire 2026 HBM capacity. CFO commentary confirmed it. Operating margin: fifty-eight per cent. Q4 revenue: thirty-two point eight trillion won. Full-year HBM revenue more than doubled year over year. The same week, Samsung’s executives confirmed a target of two hundred and fifty thousand HBM wafer starts per month by end-2026, up from approximately one hundred and seventy thousand. A forty-seven per cent capacity expansion in twelve months. In the most complex semiconductor manufacturing process humans have ever attempted.
Now layer in the demand side. NVIDIA’s Jensen Huang stated at CES 2026 that NVIDIA will be “the exclusive consumer of HBM4 in the near term.” Reports from Yonhap and UBS estimate SK Hynix supplies roughly seventy per cent of NVIDIA’s HBM4 stack volume. Samsung is racing to claim the remaining thirty per cent after passing qualification at ten to eleven gigabits per second. Broadcom CEO Hock Tan flew to Suwon to lock in supply through 2028 for Google’s TPU v8 line. Apple, finding itself unable to source LPDDR5X for the iPhone 17 and 18 lines, agreed to pay roughly double the prevailing rate to Samsung. Twelve gigabyte modules that traded around thirty dollars in Q3 2025 now ship to Cupertino at seventy.
Memory demand fulfilment rates, according to KB Securities, sit at approximately sixty per cent industry-wide. Server DRAM fulfilment is below fifty per cent. SK Hynix DRAM inventory has fallen to two weeks, from over seventeen weeks in 2024. Some DDR5 SKUs are being sold at zero inventory.
The AI boom has run into a physical wall. The wall is South Korea’s manufacturing capacity. And South Korea’s manufacturing capacity has just been hit by the most concentrated, multi-vector, single-region supply chain shock in the history of the semiconductor industry.
Three irreplaceable inputs. One geographic chokepoint. Two listed companies. Twelve weeks of buffer.
This is what is actually at stake.
III. The Story Wall Street Is Telling Itself
Pick up any sell-side note dated since the second of March. The narrative is identical across them. Iran retaliated against the US-Israeli strikes by hitting Qatar’s Ras Laffan facility on March 2nd, more comprehensively on March 18th. QatarEnergy declared force majeure. Approximately seventeen per cent of Qatari LNG capacity is offline for three to five years. Helium exports are down fourteen per cent, removing roughly eleven per cent of global helium supply. The Strait of Hormuz, where roughly twenty per cent of global oil flow transits, is effectively closed. Korean refiners are scrambling. Petrochemical crackers are running at sixty per cent. Yeochun NCC has declared force majeure. LG Chem has shut its Yeosu Number Two cracker. Naphtha has breached one thousand dollars per metric tonne. The Bank of Korea is sitting on its hands at two and a half per cent because it cannot cut into an inflation shock.
This is all true. Every number above is verifiable from primary sources, from QatarEnergy’s own X feed, from KITA trade data, from the Seoul Economic Daily, from Bank of Korea minutes, from Tom’s Hardware tracking the spot helium markets.
The consensus story stops there. The consensus story is about energy. Oil. Gas. Maybe naphtha if the analyst is sharper than average. The consensus story treats this as a 2022-style shock, painful but bounded, resolvable when the strait reopens.
The consensus is wrong about three things. It is wrong about the number of choke points. It is wrong about the location of the binding constraint. And it is most importantly wrong about whether reopening the strait actually fixes the problem.
IV. The Consensus Cannot Count
The fatal flaw in the consensus narrative is that it counts to two. Helium and oil. Maybe naphtha. The reality is that South Korea’s semiconductor industry sits on top of a triple chokehold from a single region, and beneath that triple chokehold sits a fourth, hidden vulnerability that the consensus has not yet discovered.
Choke point one: bromine. According to data published by the Korea International Trade Association, South Korea sources approximately ninety-seven and a half per cent of its bromine imports from a single country, Israel, and predominantly from a single producer at a single site. ICL Group’s Sdom facility on the southern Dead Sea extracts bromine from brine containing roughly four hundred times the bromide concentration of seawater, giving Israel a structural cost advantage that has rendered alternative sources commercially uncompetitive for the past four decades. ICL operates approximately two hundred and eighty thousand metric tonnes per year of bromine capacity, around thirty-five per cent of global production. Bromine becomes hydrogen bromide, which becomes the etch chemistry that cuts polysilicon gates in Korean DRAM. It also becomes tetrabromobisphenol A, the flame retardant that prevents AI server PCBs from catching fire under the thermal load of an eight-hundred-watt GPU. There is no semiconductor without bromine. Korea’s semiconductor industry runs on Dead Sea brine.
Choke point two: helium. KITA data again. Sixty-four point seven per cent of Korean helium imports come from Qatar. The helium is extracted as a byproduct of LNG processing at Ras Laffan. On the eighteenth of March, Iranian missiles, reportedly using complex maneuvering profiles to evade Patriot systems, struck Ras Laffan for the second time. Pearl GTL was extensively damaged. Two LNG trains were destroyed. QatarEnergy CEO Saad Sherida al-Kaabi confirmed on the twentieth of March that helium exports were down fourteen per cent and the damage will take three to five years to repair. The helium that was already in transit is largely stranded. The helium still in cryogenic ISO containers at Ras Laffan cannot leave because the Strait of Hormuz is shut and even if it weren’t shut the war risk insurance premiums make every transit economically unviable. Phil Kornbluth, the most respected independent helium analyst alive, stated publicly that helium will not be re-supplied for “a minimum of three months.” That assessment was made before the Pearl GTL damage was visible from satellite.
Choke point three: naphtha. Approximately seventy per cent of Korean crude imports come from the Middle East, almost all of it transiting Hormuz. Naphtha, the petrochemical feedstock, is what becomes propylene, which becomes propylene oxide, which becomes PGMEA, which is the solvent in which photoresist is dissolved. Photoresist is what gets exposed to ultraviolet light by an ASML scanner to define the circuits on a wafer. No PGMEA, no photolithography, no chips. Propylene oxide spot prices are up over forty-three per cent since the conflict began. PGMEA is up forty to fifty per cent. The Korean specialty chemical houses, Dongjin Semichem, ENF Technology, Dongwoo Fine-Chem, are notifying Samsung and SK Hynix of approximately twenty per cent price increases for April delivery. The procurement teams at the two memory makers are now conducting daily chemical availability checks. Daily. In an industry that runs on twelve-month forecasts.
Three choke points. Three vectors. One region. One conflict.
And then there is the fourth.
V. The Throttle Has a Name. It’s on Nasdaq.
The fourth vulnerability is the one that nobody is talking about. It is the one that turns this from a temporary shock into something potentially much worse.
Camtek Limited, ticker CAMT on Nasdaq, headquartered in Migdal HaEmek, Israel. Fiscal year 2025 revenue: four hundred and ninety-six point one million US dollars, up sixteen per cent year over year. Non-GAAP net income: one hundred and fifty-nine million dollars. Fifty per cent of that revenue, by the company’s own disclosure on the Q4 2025 earnings call, was driven by AI and HPC applications, primarily High Bandwidth Memory inspection. Camtek’s COO Ramy Langer, on the same call, stated that Camtek is “the tool of reference for 3D metrology at all major players.” Translated: Samsung uses Camtek. SK Hynix uses Camtek. Micron uses Camtek. There is no fourth alternative.
The Hawk platform, Camtek’s newest, was launched in February 2025 specifically to handle the inspection challenge of HBM4. It can inspect wafers carrying up to five hundred million micro-bumps. It can detect defects down to one hundred and fifty nanometres at twice the throughput of its predecessor. On the tenth of February 2026, Camtek announced a twenty-five million dollar Hawk order from a single tier-one IDM customer, bringing total Hawk orders from that customer to forty-five million for 2026 delivery. Langer stated, on the record, that HBM4 is “more metrology and inspection intensive” than HBM3E, because the structures are denser and the alignment tolerances tighter. Hawk and the older Eagle G5 platform, combined, accounted for approximately thirty per cent of Camtek’s 2025 revenue and are expected to exceed fifty per cent in 2026.
Nova Limited, ticker NVMI on Nasdaq, headquartered in Rehovot, Israel. Fiscal year 2025 revenue: eight hundred and eighty point six million US dollars, up thirty-one per cent year over year. Non-GAAP net income: two hundred and eighty-two point six million. Nova holds an estimated six to ten per cent of the global semiconductor metrology market, sitting third or fourth behind KLA, which holds around sixty-three per cent share, and Applied Materials. In FY2025, approximately twenty-five to thirty per cent of Nova’s product revenue came from memory customers, and memory revenue hit a record in Q3 driven by advanced DRAM and HBM. Nova’s METRION platform, an automated Secondary Ion Mass Spectrometry system, was adopted in January 2026 by two leading global semiconductor manufacturers for Gate-All-Around logic and advanced DRAM production. In March 2025, Nova and Samsung Electronics jointly received the Vladimir Ukraintsev Award for Collaborations in Metrology at the SPIE Advanced Lithography Conference. The award is specific. The collaboration is deep. The integration is years deep.
Then there is Applied Materials’ Process Diagnostics and Control business unit. Headquartered at 9 Oppenheimer Street, Rehovot, in a four thousand square metre cleanroom. The PROVision e-beam inspection system is built there. Three hundred and fifty thousand images per wafer at one and a half nanometre resolution. Israel is Applied Materials’ largest research and development centre outside the United States. Applied Materials itself published a technical brief explaining that optical metrology fundamentally cannot do what HBM requires, because the chiplets occupy too much of the reticle area, leaving no space for the optical proxy targets that conventional CD-SEM tools depend on. The transmission overlay specifications for HBM TSVs sit at approximately two nanometres. That kind of measurement requires on-chip electron beam metrology. Which means PROVision. Which means Rehovot.
KLA Corporation, the global metrology leader, acquired Israeli company Orbotech Limited in 2019 for approximately three point four billion dollars. Orbotech’s Yavne facility is now KLA’s PCB and advanced packaging inspection centre of excellence. KLA’s SPTS Technologies unit, also Israeli, makes etch and deposition tools for advanced packaging. Israel is listed in KLA’s most recent 10-K filing as one of the company’s primary research and development locations.
Stack the math. The two Korean memory giants are racing to add approximately eighty thousand HBM wafer starts per month over twelve months. Every additional wafer carries through-silicon vias whose alignment tolerance is tighter than every previous generation. Every additional wafer requires more metrology passes, not fewer. The vendors capable of providing those metrology passes for the most demanding applications are concentrated in a country at war on multiple borders, whose engineers are being mobilised for reserve duty on rotating cycles, whose ports are exposed to Houthi missile threat, whose airspace was shut for forty-eight hours during the late February strikes.
Samsung does not have a backup metrology vendor. SK Hynix does not have a backup metrology vendor. The qualification cycle for a new metrology platform on a memory product line is measured in years. The HBM ramp is measured in months.
This is the throttle. And it is being held by people whose countrymen are right now being called up to fight Hezbollah on the northern border.
VI. Four Rails, Four Speeds, One Destination
The mechanism by which this triple-plus-one chokehold transmits into the global economy runs on four parallel rails. Map them and the next twelve months become predictable.
Rail one is the petrochemical rail. Naphtha tankers cannot leave the Persian Gulf because Hormuz is shut. The shutdown is not physical. It is insurance-based. Lloyd’s of London war risk additional premiums on Hormuz transit jumped from approximately fifteen to twenty-five basis points of hull value to approximately one hundred and fifty to three hundred basis points, a four to six times increase, in the first ten days of March. For a Very Large Crude Carrier valued at one hundred and thirty-eight million dollars, that translates to indicative quotes of ten to fourteen million dollars per single transit, just for the war risk premium, on top of the cargo’s separate war risk policy. Major Protection and Indemnity Clubs, including Gard, Skuld, and NorthStandard, issued seventy-two hour cancellation notices on Persian Gulf war cover on the first of March. Hapag-Lloyd imposed war risk surcharges of up to thirty-five hundred dollars per container. The international shipping market did not need a single mine to be laid in the strait. The reinsurers closed it. By the twenty-fifth of March, Kpler tanker tracking data showed two hundred and ninety-three commodity vessel crossings since March 1st, a ninety-four per cent reduction from peacetime traffic. ADNOC CEO Sultan Al Jaber confirmed on the tenth of April: “The Strait of Hormuz is not open. Access is being restricted, conditioned and controlled.”
When the naphtha tanker doesn’t arrive, the Korean cracker cannot run. When the cracker cannot run, propylene supply collapses. When propylene supply collapses, propylene oxide jumps forty-three per cent. When propylene oxide jumps, PGMEA jumps forty to fifty per cent. When PGMEA jumps, the photoresist that Samsung’s procurement team buys from Dongjin Semichem either costs twenty per cent more in April or simply isn’t available. The photoresist shortage shows up in fab yield three to six weeks later. By June, every new wafer Samsung pushes through its lithography stack costs measurably more to produce, and a percentage of those wafers don’t make it through.
Rail two is the helium rail. Korean fabs have a supply-chain-level helium buffer of approximately four to six months. That sounds comfortable. It is not. The four-to-six-month figure is the aggregate across distributor tanks, in-transit cryogenic ISO containers, and supply-chain working inventory. The buffer at the fab itself is approximately one week. Cryogenic helium cannot be safely stored in bulk on-site because the boil-off rate is brutal and the storage infrastructure does not exist at scale. So the four-to-six-month buffer is real, but it is consumed at a rate determined by how fast the distributors can refill their tanks. Once the distributor tanks empty, the fab inventory empties within seven days. The buffer-exhaustion window, on current trajectories, falls between mid-May and late July 2026. After that, helium rationing protocols activate. Samsung’s Helium Reuse System, deployed selectively since April 2025, reduces consumption by approximately eighteen per cent at full deployment. Eighteen per cent does not bridge a sixty-five per cent supply cut.
Rail three is the bromine rail. Korean industry holds approximately six months of bromine inventory, by various estimates. Bromine is toxic and difficult to store, which caps how much can be stockpiled. ICL Group has stated in its Q4 2025 SEC filing that the security situation has not had a “material impact” on business results to date. That phrasing is doing significant work. ICL acknowledges in the same filing personnel shortages from reserve mobilisation, Houthi-related shipping disruptions, and “various bans and limitations on trade and cooperation with Israel related entities.” The risk is not that Sdom stops producing. The risk is that the bromine cannot leave Israel. Eilat is exposed. Ashdod is exposed. Air freight via Ben Gurion was shut for two days during the late February escalation. Every delay compounds.
Rail four is the metrology rail. This is the slowest-moving rail, but it is also the one that cannot be substituted at any speed that matters. Camtek, Nova, and Applied Materials Israel are not currently shipping fewer tools than they were in January. They are shipping more. Both Camtek and Nova reported record quarters through Q4 2025. The vulnerability is forward-looking. If the Iran-Israel conflict re-escalates, if a single Israeli missile defence battery fails to intercept the wrong target, if reservist mobilisation extends to the engineers who calibrate the Hawk and PROVision platforms in the field, the impact on Korean HBM yields shows up not in weeks but in months. And the impact, when it shows up, is not a price spike. It is a yield collapse. There is no spot market for HBM yields.
Four rails. Four transmission speeds. Four different categories of damage. They converge on the same two factories.
VII. Why 2019 Was Theatre
Pull the 2019 Japan-Korea trade dispute file off the shelf. In July of that year, Japan imposed export controls on three semiconductor materials destined for Korea: hydrogen fluoride, photoresists, and fluorinated polyimides. Pre-dispute, Japan supplied over ninety per cent of Korea’s high-purity hydrogen fluoride, near one hundred per cent of its photoresist, and the dominant share of fluorinated polyimides. The Korean response was immediate, governmental, and well-funded. The Moon administration committed seven point eight trillion won, approximately six billion US dollars, to materials localisation. Within three years the official statistics declared victory. Japanese hydrogen fluoride exports to Korea were down eighty-eight to ninety-seven per cent.
The official statistics lied. Or rather, they were technically true and substantively misleading. A Cambridge academic review of the post-2019 trade data, supported by RIETI and the Centre for Economic Policy Research, found that Korean photoresist imports from JSR’s Belgium subsidiary increased more than tenfold. JSR is Japanese. The technology is Japanese. The intellectual property is Japanese. Only the geographic point of last manufacture changed. The Cambridge paper concluded, and this is the line every Korean policymaker should have tattooed on the inside of their eyelids: “South Korea’s dependence on Japanese MPE products could be much greater than formally recorded.” SK Group Chairman Chey Tae-won said the quiet part out loud in July 2019: “South Korean companies are still incapable of producing hydrogen fluoride suitable for our processes.”
The 2019 lesson is not that Korea can substitute. The 2019 lesson is that substitution is theatre when the underlying technology and capacity sits in one place. The current crisis is structurally worse. Bromine cannot be substituted because the Dead Sea concentration advantage is geological, not industrial. Helium cannot be substituted because helium is a finite extracted resource, not a manufactured product. Naphtha can be substituted, in principle, but only by buying Russian crude under sanctions waivers that the Trump administration granted for thirty days starting March 12th and which expired yesterday, on the eleventh of April, with no public confirmation of renewal as of this article’s publication. The metrology tools cannot be substituted because qualifying a new metrology vendor on a memory product line takes years and Samsung does not have years.
There is also the 2022 parallel. Helium Shortage Four Point Zero. The Bureau of Land Management plant went offline. The Gazprom Amur facility had two consecutive explosions. Qatar entered scheduled maintenance. All four major helium distributors declared force majeure simultaneously. The shortage was, by 2023, declared resolved.
The 2023 declaration was wrong. The shortage was not resolved by Korean adaptation. It was resolved by demand collapse. Memory was in a brutal cyclical downturn. Helium consumption fell because chip production fell. Samsung only deployed its Helium Reuse System in April 2025, three years after the 2022 crisis. The lesson the industry actually learned was: don’t worry, the cycle bails you out.
The cycle is not bailing anyone out this time. The cycle is the opposite of what it was in 2022. The AI capex super-cycle is the largest single demand pull in semiconductor history. Hyperscaler capital expenditure is on track to exceed six hundred billion dollars in 2026, up thirty-six to forty per cent year over year. The crisis is colliding with peak demand. There is no demand-side relief valve.
VIII. The Damage Already on the Tape
What is happening, right now, in April 2026, in numbers any reader can verify against primary sources:
Korean petrochemical sector. Yeochun NCC declared force majeure on March 4th, becoming the first Korean cracker operator to do so since the conflict began. YNCC’s Number 1 cracker (915 thousand metric tonnes per year ethylene) and Number 2 (900 thousand) are both running at minimum economic load. LG Chem shut its Yeosu Number Two plant (800 thousand metric tonnes per year ethylene) on March 23rd. Lotte Chemical agreed to close its Daesan facility. The Korean government’s restructuring target reduces national ethylene capacity by two point seven to three point seven million metric tonnes per year, approximately twenty-five per cent of the national total. The largest petrochemical industry contraction in Korean history.
Korean monetary and fiscal response. The Bank of Korea held its benchmark rate at two point five per cent on the tenth of April. This was Governor Rhee Chang-yong’s seventh consecutive hold. The BOK’s February 2026 projections assumed crude oil at approximately sixty-four dollars per barrel. Spot Brent has traded above one hundred dollars throughout March and early April. March CPI hit two point two per cent year-over-year. The won breached fifteen hundred per US dollar in mid-March, the first time since 2009. The KOSPI has fallen eighteen per cent from its pre-war level, the worst four-day stretch since 2008. President Lee Jae-myung’s government activated a one hundred trillion won market stabilisation programme. MOTIR designated naphtha as a supply-chain economic security item and imposed a five-month export ban from March 16th. The 30-day Russian petroleum waiver expired yesterday. As of publication, no extension has been confirmed.
Samsung and SK Hynix posture. Both companies are publicly insisting that production is stable and inventories are sufficient. Both companies are simultaneously, according to Korean industry publication The Elec, running daily chemical procurement checks at the fab level. The internal gap between official messaging and procurement reality is the most reliable signal of impending stress in semiconductor history.
Memory pricing. TrendForce data for Q1 2026 shows DRAM contract prices up approximately ninety to ninety-five per cent quarter over quarter. DDR5 spot prices have approximately quadrupled since September 2025. Phison’s CEO confirmed that one terabyte TLC NAND chip prices rose from $4.80 in July 2025 to $10.70 in March 2026 and that “all 2026 NAND production is already sold out.” A 32 gigabyte DDR5-6000 kit on Amazon went from approximately one hundred and ten dollars to four hundred and forty-two dollars in twelve months. This is what before looks like.
Hyperscaler behaviour. According to Reuters, Google, Amazon, Microsoft, and Meta have placed open-ended orders with Samsung and SK Hynix, indicating they will accept any supply at any price. Procurement lead times have extended from twenty-five weeks to over forty-five weeks. SK Hynix DRAM inventory has fallen to two weeks. Industry DRAM demand fulfilment rate sits at approximately sixty per cent. Server DRAM is below fifty.
Downstream pain. Dell COO Jeff Clarke, on the February 26th earnings call, stated that the cost of a gigabyte of DRAM had risen approximately five and a half times in six months, to two dollars and thirty-nine cents per gigabyte. He called it “unprecedented.” HP Inc reported memory now accounts for thirty-five per cent of the cost to build a PC, up from fifteen to eighteen per cent the previous quarter. Lenovo CEO Yang Yuanqing called the situation “structural, not cyclical.” IDC revised its 2026 PC forecast from positive eight per cent growth to negative eleven point three per cent contraction. Sony confirmed PS5 disc price increases to $649.99 effective April 2nd. NVIDIA is reportedly cutting GeForce RTX gaming GPU production by thirty to forty per cent. Ford CFO Sherry House warned of approximately one billion dollars in additional 2026 costs from DRAM and inflation. Honda has cut North American volume by one hundred and ten thousand units. Smartphones face the largest projected unit decline in IDC’s recorded history, twelve point nine per cent in 2026.
This is the first-order damage. The metrology rail has not yet failed. The bromine rail has not yet failed. We are watching what happens when one and a half of the four rails are down. Imagine three.
IX. The Triage Nobody Will Admit Out Loud
Here is what almost no analyst has put in print yet. The triage is already happening. It is not a future scenario. It is current operating policy at the world’s two largest memory makers, and it is reshaping who gets to use computing in the next decade.
When materials are constrained and demand is unconstrained, the people running fabs make priority decisions. Those decisions are not made by ethical philosophers. They are made by margin spreadsheets. HBM stacks for Nvidia trade at approximately five hundred and fifty dollars per twelve-high unit, with operating margins north of fifty per cent. Server DDR5 RDIMM trades at premium pricing, with hyperscalers paying double the Q3 2025 prevailing rate to lock in supply. Consumer DDR5 trades at whatever the market will bear. Automotive DRAM trades at the lowest margins of all. Industrial mature-node memory used in defence systems sits below automotive in the pecking order.
When you run that hierarchy as a triage protocol it produces a specific outcome. Wafer capacity flows toward the highest-margin product first. Samsung executives, on background to Tom’s Hardware, confirmed this in March: “HBM and high-margin enterprise DRAM will receive priority through 2027.” Gartner’s Kanishka Chauhan said it on the record: “High Bandwidth Memory demands stronger pricing compared to traditional memory, leading vendors to allocate more production capacity to HBM.” Claus Aasholm of Semiconductor Business Review put it bluntly: “DDR4 and DDR5 are no longer strategic priorities for memory makers. They are what remains after HBM allocation.”
The hierarchy, as currently observed: HBM for AI accelerators, first. Server DDR5 and LPDDR5X for hyperscaler cloud, second. Enterprise SSDs, third. Consumer DDR5 for desktop PCs, fourth. LPDDR for premium smartphones, fifth. Automotive memory for vehicles, sixth. DDR4 and consumer NAND for everything else, last and rapidly disappearing.
DDR4 is now being end-of-lifed by all three majors. Micron has exited the Crucial consumer brand entirely, citing the need to redirect supply to “larger, strategic customers.” All three majors have issued DDR4 EOL notices. The price inversion is now visible: DDR4 trades higher per gigabyte than DDR5 because nobody is making it any more.
This produces a digital divide accelerant that nobody is talking about. Sub-five-hundred-dollar entry-level laptops, the mainstream computing platform for the developing world, run on DDR4 or low-end DDR5. They do not run on HBM. Their supply chain just got cut. Gartner’s projection: the sub-five-hundred-dollar PC segment will be “financially unviable” by 2028. IDC: “The era of bargain-priced PCs and tablets is behind us.”
In plain English: the AI boom is, right now, in real time, transferring computing capacity from poor people to data centres. Every wafer that goes into a Stargate HBM stack is a wafer that does not become a school laptop in Lagos or a delivery scooter dashboard in Hanoi or a CT scanner in a regional hospital in Bogotá. The trade-off is real. It is being made at the procurement desk in Suwon, in real time, in 2026. It has nothing to do with whether Nvidia is a good company or whether OpenAI builds beneficial AGI. It has to do with the fact that Korean fabs have one week of helium working inventory and a six-month bromine reserve and a twelve-week ramp window for HBM4 capacity, and the engineers running those fabs are choosing AI margins over consumer access because that is what their P&L tells them to choose.
There is a second-order consequence inside the second-order consequence. American defence procurement is sitting at the bottom of the hierarchy. Tomahawk missiles, F-16 avionics, radar systems, and most guided munitions use mature-node chips. Acting US Chief of Naval Operations Admiral James Kilby acknowledged on the record in early 2026: “We need to look at other vendors. They might be able to produce a missile that’s effective, which is more effective than no missile.” That is the US Navy publicly acknowledging that it cannot get the chips it needs for guided munitions because the commercial AI market has out-bid it. During an active war that the United States is participating in. The triage chooses Stargate. It always chooses Stargate. The margins are not close.
There is a third-order consequence. Chinese memory is the unintended beneficiary. CXMT, Changxin Memory Technologies, has approximately five per cent of global DRAM market share by revenue, with wafer capacity heading from three hundred thousand per month toward four hundred and fifty thousand by year end 2026. CXMT now produces DDR5-8000 and LPDDR5X-10667 with yields reportedly in the eighty to eighty-five per cent range. Nikkei Asia confirmed in February 2026 that HP and Dell have begun qualifying CXMT DRAM for non-US markets and diversified procurement respectively. The qualification cycle takes approximately six months. Which means by Q4 2026, Western PC OEMs will be shipping Chinese memory in volume for the first time, in markets where US sanctions don’t apply, because Korean memory is rationed.
The US semiconductor export control regime was designed, explicitly, to prevent this outcome. The Korean supply crisis is achieving in nine months what years of export controls failed to achieve. Iran’s missiles are doing for CXMT what Beijing could not.
Nobody planned this. Nobody is admitting it. It is happening anyway.
X. The Steel Man (And Why It’s Conditional)
The strongest counter to everything written above is that the ceasefire holds, the Strait of Hormuz reopens, the helium starts flowing again within twelve weeks, the Russian naphtha waiver gets extended, the Korean stockpiles bridge the gap, and by Q3 2026 the crisis looks like another bounded shock.
That counter-argument has weight. Let me give it the strongest version I can construct.
The ceasefire announced on the eighth of April, mediated by Pakistan, contains explicit language about reopening Hormuz. The talks scheduled for the twelfth of April in Islamabad, with US Vice President JD Vance attending, are the most credible diplomatic process to end Iran-Gulf hostilities since the JCPOA negotiations themselves. Iran has every incentive to reopen the strait because closing it costs Iran more in foregone oil revenue than it costs Iranian missiles to fire at Qatar. Qatar has no incentive to escalate further. The Trump administration has no incentive to let the AI build-out stall. The pressure from US hyperscalers on the White House to pressure Iran into reopening the strait is enormous. The structural forces all point toward de-escalation. The base case is therefore: by Q3 2026, naphtha is flowing, helium is flowing, bromine never fully stopped flowing, and the metrology vendors continue shipping at record rates because the war never reached Migdal HaEmek or Rehovot. Korea’s six-month inventories suffice. The Russian naphtha pipeline becomes a permanent backup. The triage normalises. NVIDIA gets its HBM. Apple gets its LPDDR5X at a higher price but in volume. The PC OEMs eat margin compression, raise prices ten to fifteen per cent, and the consumer absorbs it. By 2027 we are talking about this article as “the crisis that wasn’t.”
That is a coherent base case. I am not dismissing it. It might be the right base case.
Here is why it is also conditional, and what conditions it depends on.
It depends on the ceasefire holding. The ceasefire is fragile. On the same day it was announced, Israel launched a hundred Hezbollah strikes in Lebanon in ten minutes. Iran has explicitly threatened to walk if Lebanon is not included in the framework. Iranian parliament speaker Mohammad Bagher Ghalibaf alleged on the ninth of April that three clauses had already been violated. The April twelfth talks may not produce a binding agreement. Even if they do, the implementation timeline for restoring shipping insurance to pre-war levels is at minimum six to eight weeks after a credible cessation of hostilities, because reinsurers move slower than diplomats.
It depends on Ras Laffan being repairable in less than three to five years, which contradicts QatarEnergy’s own public statement on the twentieth of March.
It depends on the metrology vendors not being hit. This is the conditional that the consensus is not even pricing. If a single Iranian or Hezbollah missile reaches the wrong industrial zone in Israel, if the war re-escalates and the IDF mobilises another tranche of reservists from the technology sector, if Camtek or Nova lose three months of engineering productivity, then HBM yields fall, and the entire AI capex assumption set re-rates.
It depends on the Russian naphtha waiver being renewed, which expired yesterday and which the Trump administration has not publicly committed to renewing.
It depends on China not exploiting the moment to qualify CXMT into more Western OEMs, which it is currently doing.
The base case requires all five of those conditions to hold. The probability of any one of them failing is non-trivial. The probability of all five holding cleanly through Q3 2026 is, on my estimate, well below fifty per cent. The market is currently pricing the base case at something closer to seventy-five per cent confidence. The mispricing is in the conditional probability stack, not in the base case itself.
XI. What to Watch and When to Panic
Track these and you will know where this goes before the rest of the market does.
ICL Group Q1 2026 earnings, expected mid-May. Listen for any commentary on Sdom production, Eilat shipping, Israeli reservist impact on operations, and bromine spot pricing. Any acknowledgement of “material” impact, as opposed to the current “not material” language, is the first crack.
QatarEnergy operational updates on Ras Laffan reconstruction. Watch for satellite imagery from Planet Labs and BlackSky over the Pearl GTL site. The reconstruction timeline is the helium timeline.
KITA monthly trade data on Korean bromine and helium imports. The April figures, due in early May, will be the first month entirely under conflict conditions. If bromine imports from Israel fall below eighty per cent share, the diversification has begun. If helium imports from non-Qatar sources exceed fifty per cent share, the rerouting is working.
Naphtha CFR Japan benchmark. Currently trading above one thousand dollars per metric tonne. A return below seven hundred signals supply normalisation. A breach above twelve hundred signals further escalation.
Korean petrochemical operating rates. Watch for any force majeure rescissions from YNCC and LG Chem. Watch for whether Lotte Chemical and Hanwha Total declare new force majeures.
Samsung and SK Hynix Q1 2026 earnings. Samsung expected late April, SK Hynix in May. Listen for any downward revision to HBM4 capacity guidance. CFOs do not volunteer this kind of information unless they have to.
DRAM and NAND spot prices on DRAMeXchange. DDR5 spot above one thousand dollars per module signals ongoing crisis.
Camtek and Nova quarterly earnings. Camtek expected in early May, Nova in mid-May. Watch for any acknowledgement of operational disruption from the Iran-Israel conflict.
MOTIR emergency announcements. Any expansion of the 14-material list, any new export bans, any new strategic stockpile releases.
Strait of Hormuz transit count via Windward Maritime AI and Kpler. Currently at five to seven daily transits versus a normal seventy-seven. A return above thirty per day signals meaningful reopening.
Iran ceasefire negotiations in Islamabad. April 12th talks are the binary inflection point.
Lloyd’s war risk additional premium quotes for Hormuz transit. A return below fifty basis points signals insurance market normalisation, which is a necessary precondition for shipping flows to recover.
US Treasury Russian petroleum sanctions waiver status. Expired yesterday. A renewal announcement is a structural positive. Silence is a structural negative.
CXMT qualification announcements at Western PC OEMs. Each qualification is a permanent loss of Korean market share that does not come back when the crisis ends.
XII. The Throttle Has Always Been There. We Just Didn’t Look.
The story everyone is telling is that the Strait of Hormuz closed and the world’s energy markets shuddered. That story is true. It is also small.
The story nobody is telling is that two factories in a country at war on three borders make the inspection tools without which Korea cannot ship the chips that power the AI boom that the entire global economy is now structurally dependent on. The throttle has always been there. We just didn’t notice it because nobody had any reason to look.
The Hormuz closure forced everyone to look. And what we are seeing, when we look properly, is that the AI super-cycle is balanced on a much narrower platform than its valuations suggest. Three irreplaceable inputs from one region. Two companies in one country handling the metrology. Six months of bromine. Four months of helium at the supply-chain level, one week at the fab. Twelve weeks until the buffer-exhaustion window opens. A ceasefire whose terms Iran’s parliament is already alleging have been violated. A Russian naphtha waiver that expired yesterday. A triage protocol that is sacrificing the consumer electronics market and the Pentagon’s munitions supply chain to protect Nvidia’s HBM order book.
The molecules are trapped at Ras Laffan. The bromine is stranded at Sdom. The metrology engineers are on rotating reserve duty. The Korean procurement teams are running daily checks. The wafers are still being made. The HBM is still being shipped. The throttle has not yet closed.
But the throttle has never been more visible. And the people who control it have never been less in control of their own circumstances.
This is not a bounded shock. This is a structural rebalancing of who gets to use computing in the second half of the 2020s, decided in real time, by procurement managers in Suwon and underwriters in London and missile operators in Iran, none of whom are coordinating, all of whom are reacting, and the answer they are converging on is the same: AI first, everything else second.
Watch the Islamabad talks on April twelfth. Watch the Camtek earnings call in early May. Watch the Korean April KITA data. Watch the Lloyd’s war risk premiums.
And remember, when the next quote-tweet asks why the AI build-out is suddenly stalling in Q3 2026, that the answer was already visible in April, in the gap between what Samsung was telling its shareholders and what its procurement team was doing every day at 9 AM in the Suwon supply room.
Two Israeli companies. Three choke points. One closed strait. Twelve weeks of buffer. Seventy per cent of NVIDIA’s HBM4 line. Forty per cent of global DRAM committed to a single customer. Six hundred billion dollars of hyperscaler capex riding on it.
This is the throttle. Now you can see it.
Veron Wickramasinghe writes on geopolitics, energy, and the upstream supply chains that shape global power. The Throttle is part of an ongoing series on the structural vulnerabilities exposed by the 2026 Hormuz crisis. Earlier pieces in the series include the Qatar helium analysis, the Methane Trilemma, the Sulfur Trap, and the AI Supply Chain analysis. Follow @veronken for the next installment.
Sources
All quantitative claims in this piece are verified against primary or Tier 1 sources, including:
Trade and commodity data: Korea International Trade Association (KITA) bromine and helium import data for 2025; United States Geological Survey Mineral Commodity Summaries 2026 (helium, bromine, iodine); Korea Customs Service monthly trade releases; Kpler tanker tracking data on Strait of Hormuz transits (March 2026); Windward Maritime AI vessel transit reports.
Company filings and earnings: Camtek Limited (NASDAQ: CAMT) Q4 2025 earnings call transcript and SEC Form 6-K filings of February 10 and February 18, 2026; Nova Limited (NASDAQ: NVMI) Q3 and Q4 2025 earnings releases and Form 20-F annual filing; ICL Group Limited (NYSE: ICL) Q4 2025 results release of February 17, 2026 and SEC Form 6-K filings; SK Hynix Q4 2025 earnings release and CFO commentary; Samsung Electronics investor relations disclosures (December 2025 through March 2026); Applied Materials Process Diagnostics and Control technical brief on HBM eBeam metrology; KLA Corporation FY2025 10-K filing.
Government and institutional sources: Bank of Korea Monetary Policy Board statement of April 10, 2026 and February 2026 economic projections; Korean Ministry of Trade, Industry and Resources (MOTIR, formerly MOTIE) press releases on the 14-material emergency review and the March 16 naphtha export ban; QatarEnergy official statements and CEO Saad Sherida al-Kaabi press conference of March 20, 2026; United States Treasury Department Russian petroleum sanctions waiver of March 12, 2026; International Energy Agency strategic reserve coordinated release announcement of April 7, 2026; United States Securities and Exchange Commission filings; Lloyd’s Market Association war risk circulars of March 2026.
Tier 2 expert and journalism sources: Carnegie Endowment for International Peace (”The Iran War Is Also Now a Semiconductor Problem,” March 2026); Center for Strategic and International Studies (”The Impact of the Iran Conflict on South Korea: By the Numbers,” March 2026); Reuters and Bloomberg coverage of the OpenAI Stargate Letters of Intent signing in Seoul (October 2025); Nikkei Asia reporting on CXMT qualification at HP and Dell (February 2026); Financial Times and Wall Street Journal coverage of the Ras Laffan strikes; TrendForce DRAM and HBM contract pricing analyses (December 2025 through April 2026); Gartner, IDC, and Counterpoint Research market projections; Tom’s Hardware coverage of the Korean memory triage and the helium two-week clock; The Elec (Korea Electronics Industry Media) reporting on Korean fab procurement protocols of March 27, 2026; Seoul Economic Daily, Korea Economic Daily (KED Global), Korea Herald, BusinessKorea, and Yonhap News for Korean petrochemical and policy coverage; Hydrocarbon Processing and ICIS for cracker operating rate data; gasworld for helium spot pricing and inventory tracking; S&P Global Commodity Insights, Argus Media, and OPIS for naphtha and crude pricing; Phil Kornbluth (Kornbluth Helium Consulting) for industry expert commentary; the Cambridge Core / RIETI / Centre for Economic Policy Research academic review of the 2019 Japan-Korea trade dispute outcomes; Al Jazeera, Anadolu Agency, NPR, and Time for conflict timeline reporting; Mordor Intelligence for Israeli semiconductor sector reservist mobilisation data.
Earnings call and executive commentary citations: Dell Technologies February 26, 2026 earnings call (Jeff Clarke); HP Inc. earnings disclosures (Karen Parkhill); Lenovo Group commentary (Yang Yuanqing); Ford Motor Company earnings call (Sherry House); Sony Group PS5 pricing announcement of April 2, 2026; Acting United States Chief of Naval Operations Admiral James Kilby on the record commentary on defence chip procurement; SK Group Chairman Chey Tae-won at GTC 2026; NVIDIA CEO Jensen Huang at CES 2026; Camtek COO Ramy Langer Q4 2025 earnings call.
Errors of interpretation are the author’s alone.