A reading desk, not a publisher.

Rackline pulls public RSS and Atom feeds from specialist outlets and national technology desks. It keeps stories about data centers, power, chips, models, infrastructure technology, policy, and the capital moving through them, then lays that edition out like a newspaper.

The front page is the United States edition. Stories that never name a country stay there. A story that names somewhere else leads only when it is a blockbuster: a frontier-model launch, a tens-of-billions deal, export controls, or a multi-gigawatt build. The rest are filed under Europe, Asia-Pacific, the Middle East and Africa, or the Americas outside the United States. A story can sit on more than one desk in the file, and a quiet desk borrows a fresh neighbour so the desk page still has something to read. The front prints that story once. A desk or a rail with nothing in the day, and nothing large enough from the week, is left off the page. The desks themselves do not change.

Newsroom copy files through POST /api/desk with a bearer key. The article is the body. The link is the document it reports. A company report or an earnings update for a name in the earnings book is filed onto that block. The block keeps a report link beside the chart, and the link opens the report in Rackline. Anything else is a desk filing. A requested desk is ignored. When an OpenAI key is set, that model reads the filing before it is stored. It sends the piece back when the copy is off the beat, promotional, self-contradictory, or contradicted by the linked document, and it assigns the desk. It does not rewrite the article. Without that key, keyword rules assign the desk, and off-beat, promotional, and thin leads do not file. A headline a specialist outlet already holds keeps that outlet’s slot. The reader shows the filed paragraphs.

The page is ordered on freshness and consequence. Copy is banded by age, and a story normally stays behind anything in a newer band. A stated commitment of at least $1 billion or 300 MW may bridge one six-hour band, so a consequential filing does not fall beneath a small update posted minutes later. It cannot bridge the day. Inside a band, explicit United States filings rank ahead of unlocated copy, then stated money and megawatts, named counterparties and instruments decide the order. Campaign reads, gadget stunts, and broadcast segments count against. An ordinary office letting stays off: a tenant that describes its product as artificial intelligence has not leased a hall, a campus, or the power. The wire keeps 45 days so the deals book can read 45 days. Ordinary news leaves the front, the desks, the regions and search after 24 hours. A foreign filing reaches the United States edition only as a blockbuster, or on its own regional rail while it is still news.

Ranking also asks who a filing is for. A story scores higher for the size of the figure it states, for naming parties that can actually commit capital, power, or silicon, and for naming the instrument — a tranche, a facility, an offtake, a prelease. It scores lower for the things a trade press writes when there is no news: round-ups, listicles, predictions, buying advice, survey findings. Rackline is written for people who sign things, and the order of the page is the main place that shows.

Every desk runs long when its own file is deep enough. The front gives each desk four filings, then four more that the desk itself filed, after the brief and the regions have taken their slots. Power carries campuses, large loads, interconnection, and the generation built for them. Chips carries packaging, memory, foundries, and the rack silicon. Technology carries the hall: cooling, 800 VDC, and the rack. Policy carries permits, moratoriums, rate cases, and export rules, including the statehouses where the siting fights actually happen. Capital carries the money in those. Models & Labs treats agents, model serving, inference platforms, evaluation, retrieval, fine-tuning, observability, guardrails and model-support tooling as part of the model industry—not as generic enterprise software. Major launches from foreign frontier labs can reach the United States edition; ordinary foreign coverage stays on its regional rail.

Feeds only publish what is on their own front page, which on this beat is a fortnight on a good day. So each refresh folds its pull into a rolling 45-day archive rather than replacing the one before it, keeping the fuller telling where the same item comes back richer and pruning whatever has aged out. The archive deepens the longer the site runs and survives a restart through a snapshot on disk. It never fetches anything itself; it only remembers what the wire already pulled.

One event takes one slot. When several outlets file the same raise, contract, or ruling, Rackline groups the retellings — the figure, the company the headline leads with, and the thing that happened, so a White House sitting, a buyback, a prospectus, or a product launch written four ways is one story — and runs the telling that carries the article. A shared company name does not glue two filings together, and a side mention of someone else’s round does not become that round. A summary does not take the slot when a full reprint of the same event is already on the wire. Among full reprints, the best-sourced outlet leads, and that is the one the desk runs. A second telling does not scroll past it. A company that files all day does not take the next several slots. The desk opens on that news, then reads someone else when someone else has filed, including a large filing from earlier in the week. The rest of that company’s day is still on the desk. A company that led yesterday does not lead the front again when another company has filed. That story can still run once, lower on the page.

A headline opens the article in Rackline. The page is the piece, set in the reader and credited to the publisher. When the feed already carries the whole article, that text is the reading. A short feed summary is not the piece. The publisher page is read, headings and the later paragraphs stay, a table in the piece is kept as lines, and the longer body is the reprint. A comment form, a related list, and a newsletter block are not the article. A line that only says to continue elsewhere is followed once when that link is the rest of the article. Share opens the phone sheet when the browser has one, and otherwise copies the link. Download PDF writes the headline, the credit, and the paragraphs on the page into a file. The page does not explain why the item was filed, and it does not grade the source. On a phone, a short downward pull from the top reveals a compact Back control that exits the whole reading sequence; it stays collapsed while articles are swiped. A swipe to the right retraces articles already crossed; from the originally opened article, one more returns to its page and position. A swipe to the left advances to the next filing. That order is the page the story was opened from: the front, from the banner through the brief, the desks, and the regions, or one desk from the top of its file to the bottom. A story is in that sequence once. At the end of a desk, the swipe opens the next desk. The warmer has already read those articles, and the phone prefetches the next few in full, so the swipe is the piece already in hand rather than a page still loading. A $ next to the source marks a publisher that keeps the original behind a paywall. Those stories stay on the publisher’s site. Rackline does not fetch them. When another outlet has published the same filing in the open, that reading is used and credited.

Each post to the desk is read by OpenAI when that key is set, before the piece is stored. That read can send the piece back or assign its desk. It does not rewrite the article. Page views do not call a model. The warmer may ask OpenAI for one public address of a paywalled filing. It does not ask the model to write the article. Separately, when an Anthropic or OpenAI key is set, an editor rewrites the edition once every 15 minutes for the whole site. It picks the lead and the brief, moves stories filed on the wrong desk or region, drops promotional and duplicate items, and writes a short What changed note. It sees only headlines and feed summaries, never the articles, and every point in the note links the stories it draws on. Claude is preferred for that edition pass when both providers are set. The desk filing is OpenAI only. Without a key, or if the configured models do not answer, the rules below set the page and keyword scoring assigns a filing.

Under the credit, before the photograph, a Useful control offers thumbs-up or thumbs-down beside share and download. The mark is kept with the headline, the source, and the desk. One reader counts once on a story. A second tap changes or clears that mark. A new story marked again inside a second and a half, or a run of marks from one reader or one network in an hour, is not counted. Rackline keeps the time and a hash of the network address for a week so a burst can be reviewed. It does not keep the address itself, and it does not send that record to the model. Once a day, when the OpenAI key is set and the week has marks from several readers, that key writes a short standing note: what kind of filing to surface, and how to order the page. The edition editor and the desk editor read that note. The edition editor sees a story’s count only after three readers have marked it. Neither one rewrites the article. A page view does not call the key. Without the key, or without that many readers, the marks are kept and the rules set the page. A single mark does not move the lead.

Indexes combines live GPU rental asks, gas, wholesale power, facility lease rates, vacancy, absorption, construction, and hyperscaler capex. Units stay separate except for Gas-to-Chip, which is the one deliberate conversion on the page and is described in full below. Rackline’s pressure indexes publish their weights and normalization; brokerage inventories are never added together. The AI market view may interpret those supplied readings but cannot invent a figure. A figure on the indexes page opens the release, the exchange, the ISO, or the brokerage book it came from. A Rackline ratio, including the pressure indexes and the Gas-to-Chip conversion, opens the input it was built from, or this page where the arithmetic is Rackline’s. Episodes, on the banner after the Americas, is a listening list of podcasts and videos from the last three weeks. Every 15 minutes, the warmer reads an approved set of publisher podcast feeds and keeps episodes on power, compute, grids, chips, capital, and the buildout. TBPN is a standing show: the latest full episodes stay on the list, and the short Diet recut of the same day does not. A feed that omits an episode page still counts when it carries the publisher audio. A publisher’s RSS enclosure is the first choice for listening inside Rackline. A newly seen audio address is checked before it is kept, and an accepted player is reused on later refreshes. When a feed has no playback address and an OpenAI key is set, the warmer may search for the official audio or video file, or a publisher player; it does not generate or transcribe the episode. Direct publisher video and HLS streams play in the same control as podcast audio. The title opens the Rackline listening page. The share icon copies that address. The publisher stays linked from the episode. When the publisher posts a transcript, the warmer keeps that text. When an OpenAI key is set, it writes a short desk note from the transcript. Rackline does not transcribe the audio itself. Page views call neither the feeds nor the model, and a feed failure leaves the last good edition in place. An accepted player is carried into later refreshes, and a new item does not enter Episodes until the warmer has a validated in-Rackline player for it. A guest’s figure stays the guest’s.

The plant table on Indexes is the rack a hall has to accept, taken from three NVIDIA notes and not from a model. GB200 NVL72 shipped in 2024 and GB300 NVL72 in 2025, both 72 GPUs on the MGX footprint. Vera Rubin NVL72 ships in the second half of 2026, with 72 Rubin GPUs, a 45°C liquid inlet, and a liquid-cooled busbar to 5,000 A. An 800 VDC power rack in that same half feeds the row from the building’s existing AC. A row bus of up to 2 MW is expected in 2027. Kyber, dated 2027 in NVIDIA’s OCP note, puts 144 GPUs in one rack and moves in-rack parts from 54 VDC to 800 VDC. NVL576 is eight racks. Feynman on Kyber is an NVL1152 scale-up, and NVIDIA has not dated it. The kilowatt column is a different NVIDIA document from those three notes. The DGX GB rack guide states about 120 kW of rack consumption. The GB300 NVL72 reference states a full rack of up to 142 kW. The DSX facilities guide, v2.0 of 19 August 2026, states a 330 kW cabinet TDP for Vera Rubin NVL72 and an MGX Gen 1.1 cabinet at 198 kW. Each print stays in the words its source used. A megawatt of that stated rack power is 8.3 GB200 racks, 7.0 GB300 racks, or 3.0 Vera Rubin racks. Kyber and Feynman have no stated kilowatt, and the 2 MW figure remains the row bus. The same DSX guide sizes an example site at 972,500 square feet of building for a 250 MW-class IT load: 3,890 square feet per megawatt, or 2.57 megawatts per 10,000 square feet of that building. The Technology desk is the long reading of that plant: cooling, the silicon in the rack, the 800 VDC architecture, and rack density, with this same table. A cold plate, a CDU, immersion, or a direct-to-chip loop files there. A cooling plant that belongs to a campus stays on Power & Sites. The other desks remain the day’s file.

The Apple and Android app reads this same edition. Front, Indexes, and Search call /api/mobile, and a headline opens /api/mobile/story. Those routes use the page reads. They do not call a model, and they do not add a figure that is not already on the desk. The story screen is the article. On a phone, a right swipe retraces the reading sequence and exits only after the originally opened article, while a left swipe advances in that page’s reading order. At the end of a desk, the left swipe opens the next desk. A $ marks a paywalled publisher.

Earnings sits after Episodes. It is one hundred latest quarters, most pertinent to a data-center book first, and the page is searchable by company, ticker, or cohort. The first nine are the release notes for Nvidia, the buyers, Oracle, CoreWeave, GE Vernova, and Constellation. Each of those cards is the print and one paragraph on what that print means for offtake, credit, chips, or a power date. The figures are the company’s, taken from the release, and each one opens that release. A grey bar is an outlook the company gave, not a result. When an OpenAI key is set, those paragraphs ask OpenAI first. A sentence that brings a figure of its own is dropped. Every name after those nine is the same block, through the last company in the book: company revenue this quarter and a year earlier, diluted earnings per share when that quarter states one, and the reason the name is on the desk. Those figures are rounded from the latest quarterly SEC filing, not a data-center segment. A year-ago figure is the same line a year earlier, when the filing states it. A page reads the last warm copy and the saved book. It does not call a model, and it does not call the SEC. Each block keeps a twelve-month price chart closed. Opening it loads that name’s daily bars from TradingView. Rackline does not store the price.

Three more readings sit on the indexes page and are not folded into the tape. Moratorium Nation counts local pauses on data centers, batteries, solar, wind, and crypto. AI GridWatch counts data-center actions only, including zoning fights. The two counts are printed side by side and are not added. New York’s permit hold and Texas’s grid and permit pauses are state actions, listed under the books rather than inside them. The credit health check prints one instrument, Oracle’s five-year CDS, on a display ceiling that is a drawing scale and not a rating, plus a by-firm table of published marks. Cash capex on that table is labeled as cash capex. When an OpenAI key is set, this is the reading that asks OpenAI first, along with Earnings; Anthropic is the fallback, and a sentence that brings a figure of its own is dropped. Without a key, the sentences are written from the locked marks. The extinction meter is the median answer, from the February 2026 Summit on Existential Security, to the probability of human extinction or permanent disempowerment before 2100. The respondents already work on that question, and the page says so. Beside it are the Future of Life Institute’s Summer 2026 grades. The meter is not a Rackline probability. Under that meter, launch cost is dollars per kilogram to low Earth orbit. Orbital Radar and SpaceNexus do not print the same Falcon 9 figure, so both stay: about $3,246 per kilogram from a $74 million list over 22,800 kilograms, and about $3,000, or about $2,700 reused, from Orbital Radar’s list. Falcon Heavy is about $1,520 per kilogram on SpaceNexus alone. Electron is about $25,000. The Shuttle-era print is about $54,000. Rideshare is a different product, about $6,000 on one desk and about $7,000 on the other. Starship has no commercial price. The design target and the analyst band are labeled as targets. Rackline does not invent a hall’s mass or a year when the rent crosses over.

The market matrix runs on one house, CBRE, so the megawatts are comparable down the column. Two of its figures are arithmetic rather than prints: vacancy is available capacity over inventory, and share of book is inventory over the eight-market total. Both are computed from the same published table, and the vacancy column is checked against the rates CBRE states separately. Under construction is the megawatt count when CBRE printed one, with the change or the prelease share under it, and the constraint CBRE published when it did not print a count. The dollar column is the only primary-market ask on the index, in dollars per kilowatt-month. It is a 2026 ask when a house printed one this year, with the house, the requirement, and the period under the figure. H1 2026 printed that range for Northern Virginia at 10 MW and up, and for Silicon Valley at 250 kW and up. Chicago’s 2026 dollar is still CBRE’s Q1 band at 250–500 kW, because the half printed a move and not a new range. Where the half published a move and not a dollar, the cell says there is no 2026 dollar. Atlanta, Dallas–Fort Worth, and Phoenix keep JLL’s H1 2025 cards at 1–5 MW underneath, as the last range. Hillsboro’s last range is CBRE’s year-end 2025 release, and that sentence does not name a size. Phoenix’s H1 2025 card is later than CBRE’s Q1 2025 average, and it is still not a 2026 dollar. New York Tri-State carries JLL’s New York card and New Jersey card underneath, not a blended range. An older print of the same market is not listed again as a second rate. Pacific Northwest is not Hillsboro, and Northern California is not Silicon Valley, so those cards stay off the eight-market book. H1 2026’s chart of 250 kW-and-up asks was not published as text, and a percent is not turned back into a dollar to fill it. The lease signal stays the H1 move. Where that move is a rate, it is the 10 MW+ tier from the same book. The column does not replace the national size curve from JLL’s midyear 2026 book. Cushman & Wakefield’s H1 2026 update and Colliers’ marketplace print capacity. Data Center Dynamics’ recent pieces quote percent moves or older national averages, not a 2025–26 hub range. The last primary-market average CBRE printed, $195.94 per kW-month for a 250–500 kW ask in H2 2025, stays the eight-market average. It is not printed under the matrix. A percent is not turned back into a rate.

The capex book is the demand side of the same market. It carries quarterly purchases of property and equipment for Microsoft, Alphabet, Amazon, and Meta, taken from each company’s own cash-flow statement and placed on the calendar quarter it closed. Rackline does not estimate a data-centre share of those dollars, because none of the four discloses one; the figure is corporate capex and is labelled as such wherever it appears. Full-year guidance sits in its own table and is not added to the quarterly one, because two of the four guide a number that includes finance leases and two do not. The combined column is Rackline’s sum of the four filings rather than a published total.

Capex Momentum is the third pressure index. It weights the sequential move in combined quarterly capex, the move on the same quarter a year earlier, how far the guided year has been re-marked, and capex as a share of operating cash flow. That last component is the one a lender reads: a cohort spending more than it earns is funding the next hall in the market rather than out of the business.

The masthead names the hour in New York: Morning from 5 a.m., Afternoon from noon, Evening from 5 p.m., and Night from 9 p.m.

Deals & Risk is one book, not two. It ranks the last 45 days on three axes with published weights: size at 0.45, scope at 0.20, consequence at 0.35. Size is read from the headline and nowhere else, so the figure on a row is the figure the publisher put on it and a number buried in the body of a feature cannot promote that feature onto the book. Money and capacity are scored on a log scale, because the gap between $200m and $2bn matters more than the gap between $20bn and $22bn, and a currency is never converted — a sterling row prints sterling and stays out of the dollar total. Scope counts the named parties, how many distinct roles they occupy, and how many desks and datelines the filing touches. Consequence is what an instrument of that type does to the market behind it, plus how many houses filed it, plus whether the copy calls it a first or a record.

The book reads the verb before it reads the number. Signed means the copy says it happened and does not hedge. Reported means it says it happened and hedges anyway, which is how a scoop reads before the filing lands; those rows are discounted and kept in a rail beside the book rather than in it. A target with a year on it, a tender, an exclusive-talks story, an offer, a plan, an “up to” — none of these is a signed deal. When several houses file one event and only some of them print the figure, the row uses a telling that states the figure and names a party the lead already named, so a paraphrase that says only “multibillion” does not hide the number. A forecast, a revenue print, an investor-day target, a market slide, and an insider stock sale are not transactions, however large the number. Round-up columns and analysis pieces are excluded outright, because the figure in a column belongs to whichever item the column led with and often is not on this beat at all. To be material a transaction needs nine figures, a block of capacity, or, where the sum is undisclosed, two named parties on a kind of deal that moves who owns or controls capacity.

Counterparties come from a finite register of parties that can commit capital, power, or silicon, checked by name. Rackline would rather say a party is not in the register than guess that a capitalised word is a company, so a row whose parties it cannot place is placed by desk and dateline instead. Trends on that page are measured only against filings that cleared the deals floor, so a month of commentary about private credit does not register and two priced tranches do. A theme needs two transactions and at least one of them signed. A risk is a different mark on the same list. The headline has to state a crack that has happened — a force-majeure notice, a delayed payment, a downgrade, a default, a halt, or a customer pulling out of a campus already committed — on a named party or a named project. That row can outrank a smaller signing. It is not added to the signed-capital total. An essay that only raises concerns is still not a row.

Pages do not compile the wire. A warmer pulls the feeds, the commodity tape, the research list on a schedule and stores the result. A page reads that result. The first visitor after a restart is served the last snapshot while the next pull runs, instead of waiting on ninety feeds.

Breaking is a single slot. A filing from the last 12 hours can be considered when it is on the order of a billion dollars or a gigawatt. Before the banner runs, the desk asks the OpenAI key whether that filing would move a financing or a build. A column, an opinion, or a note on what a smaller sum implies stays off the banner. When the key is absent or the model declines, the slot stays empty. After a banner, the slot stays quiet for a week unless a clearly larger filing arrives and the model agrees. Most days there is no breaking line.

Research reads arXiv’s latest lists in distributed computing, hardware architecture, and systems, and keeps a paper only when the title is about compute, power, or the plant. It reads patents the same way: the title has to name a data center, and the publication has to be inside eighteen months. A search hit that merely contains the words does not get a row.

Feeds and the commodity tape are cached for 15 minutes and pulled again by a background warmer. The open edition also redraws itself every two minutes from that cache. The Associated Press, Reuters, and Bloomberg license their wire. They do not publish an open feed Rackline can rewrite, so their reporting shows up here only when another source on the list covers the same event.

The strip under the masthead is the physical book behind a data center: rental GPUs, Henry Hub gas, and power in the main data-center markets. It scrolls on a phone the same way it scrolls on a desk. There are no arrows. A finger or a pointer can slide it while it is moving, and it picks the crawl back up when you let go. A reader who has asked the browser for less motion gets a still strip they can still slide.

Compute is an on-demand rental print in dollars per GPU-hour for the A100 80GB, H100 SXM, H200, B200, and B300. Rackline reads the per-provider asks compiled by Price of Compute. Each provider gets one vote, the median of that house’s firm on-demand asks. A second region from the same house folds into that vote. The book is firm on-demand asks. Once five houses are quoting, asks outside the 10th to 90th percentile are set aside, and the print is the equal-weight average of what remains. Public books do not say how many GPU-hours are actually offered, so that leg is not weighted by capacity. Fewer than three surviving votes is marked as a thin tape.

The cell on the tape is one price per chip, Rackline’s weighted index. Each house still in the ask book counts as weight 1, so that leg enters as the ask mean times the number of houses. Ornn’s OCPI settlement counts as weight 1. CoreWeave’s card — the published node price divided by the GPUs in the node — counts as weight 1, and that house is then left out of the ask book so it is not counted twice. A node with no on-demand price, which is how CoreWeave lists the B300, is left off. A leg that did not print is left out, not filled in. CME and ICE have announced GPU futures and have not posted a price, so neither is a leg. The index is a rental. NVIDIA does not publish a price to buy the board, and no OEM quote or filing is on this tape, so Rackline does not print one. Gas-to-Chip uses this index.

Henry Hub is the NYMEX front-month natural gas future, in dollars per MMBtu, read from Yahoo Finance. The change is against the previous close on that chart.

Gas-to-Chip

Gas-to-Chip asks one question: what is a MMBtu of natural gas worth once it has been burned, cooled, and sold as an hour on a frontier GPU? The answer is published in dollars of GPU revenue per MMBtu. The GPU rung is the weighted index above. The other rungs are published prints.

1. Henry Hub, front-month, off the tape above.
2. Heat rate. 7.2 MMBtu/MWh, the figure EIA reports for new combined-cycle gas. One MMBtu therefore buys 138.9 kilowatt-hours at the generator busbar, and the fuel-only generation cost is Henry Hub × 7.2 dollars per MWh.
3. PUE. 1.30, so 106.8 of those 138.9 kilowatt-hours reach a GPU and the rest is cooling and distribution loss. The same 1.30 puts the fuel bill on an IT-load basis, which at 730 hours a month is what a wholesale lease is quoted against.
4. The market lease ask. Not the hyperscale rung on its own. Every band on JLL’s North American curve is weighted by the megawatts it represents — at the band midpoint, and at the floor for the open over-20 MW band, because JLL publishes no ceiling and Rackline will not invent one. Retail halls ask the most per kilowatt and carry well under one percent of the capacity, so the capacity-weighted market ask lands near $161/kW-month against $152 for large blocks. Adding the fuel gives the all-in cost of a powered kilowatt.
5. The GPU index. The weighted print above, divided by board TDP, which turns dollars per GPU-hour into dollars per IT kilowatt-hour. Multiplied by the 106.8 IT kilowatt-hours a MMBtu delivers, that is GPU revenue per MMBtu.

Each part’s line prints that weighted index in dollars per GPU-hour. The figure at the right is the same index after the watt conversion, in dollars of GPU revenue per MMBtu. B300 is the 1,400 W liquid-cooled bin. Its on-demand CoreWeave card is blank, and Ornn does not settle it, so the hourly figure is the ask mean alone. The headline runs on whichever part converts a MMBtu into the most revenue, which is a question of ask per watt rather than ask per hour: a 1,400 W board asking more per hour than a 1,000 W board can still be the worse converter, and currently is. The gas multiple compares that revenue with the gas price alone. The all-in multiple charges the fuel and the powered space together, and is the more honest of the two.

What Gas-to-Chip is not: a delivered power price, a contracted rate, or a margin. It carries no capacity payment, no basis, no transmission, no capital cost, no network, no staff, and no utilization haircut. It is a conversion between two live markets, and its only real claim is about proportion — fuel is a fraction of a percent of what an hour on a frontier GPU sells for, so gas is not what makes compute expensive.

The real-time power index is the equal-weight mean of the live ISO hubs, in dollars per megawatt-hour. A hub that did not answer is left out. Henry Hub, the daily peak trades, and AEP Ohio’s price to compare are printed under it and are not votes in it. Real-time cells are the latest posted interval. Peak cells are the newest daily on-peak trade in the EIA wholesale file, and only if that trade is from the last three weeks. The trade date is printed on the cell.

• ERCOT North, real time. Dallas and the rest of the North Texas load.
• ERCOT West, real time. The Abilene corridor.
• ERCOT Houston, real time. The Gulf Coast hub.
• CAISO NP15, real time. Silicon Valley and the Bay Area.
• CAISO SP15, real time. Los Angeles and the Inland Empire.
• MISO Illinois, real time. Central Illinois. Chicago itself is the ComEd zone.
• MISO Indiana, real time. The liquid Midwest hub, and a lot of the Indiana build.
• Columbus and Waterford, the AEP Ohio generation price to compare. Both towns are on AEP Ohio, so they share one published rate. It is the standard-offer generation price, not the PJM nodal price a campus pays.
• NYC Zone J, real time. New York City.
• Long Island, real time. The NYISO LONGIL zone.
• Palo Verde, daily peak. The hub Phoenix buys against.
• Mid-Columbia, daily peak. Hillsboro, Quincy, and the rest of the Columbia River load.
• Mass Hub, daily peak. New England, including the Boston-area campuses.
• PJM Western Hub, daily peak. The liquid eastern hub. Northern Virginia is the Dominion zone, and that zone price is not this print.

Atlanta sits on Southern Company. That utility does not publish a hub in this public set, so Atlanta is not on the tape.