BUSINESS

Some contract announcements move stocks. And then there are the ones that reframe an entire investment thesis. We saw the latter on a Monday, July 20. IREN Limited (IREN) was trading up more than 21% to $40.57 as of July 20 midday, according to Yahoo Finance.That’s after the Australia-based Artificial Intelligence (AI) cloud infrastructure company announced it had signed $2.8 billion in new multi-year cloud services contracts and raised its year-end AI Cloud annualized run-rate revenue target to more than $4 billion, up from $3.7 billion, according to IREN’s announcement.What makes this announcement more than a headline is who those contracts are with. The company’s customer base now includes Microsoft, Nvidia, Perplexity, Figure AI, Together AI, Fluidstack, Fireworks AI, Fal AI, and Hume AI, plus a new undisclosed leading AI developer. When Microsoft and Nvidia are in your customer list alongside the fastest-growing AI application companies in the world, of course, the market pays attention.Also Read: IREN Limited Latest News and StoriesWhat IREN’s $2.8B in contracts actually representsApproximately 85% of IREN’s $4 billion-plus ARR target is now under contract, according to IREN’s disclosure. That is not a pipeline figure. That is committed revenue with signed agreements. The weighted-average contract term across IREN’s portfolio is approximately four years, providing revenue visibility well into 2029 and 2030.More IREN Stories:Howard Lutnick’s Cantor buys $126 million of surprising AI stockFast-growing AI player taps markets for billions amid data-center frenzyPopular analyst unveils two new AI stocks for 2026Two details inside the announcement stand out to me as particularly meaningful for investors.First, recent contracts include customer prepayments representing approximately 45% of the associated GPU capital expenditure. In plain terms, IREN’s customers are funding a significant portion of the infrastructure buildout themselves, reducing the company’s net capital requirement for those deployments. That dynamic meaningfully changes the risk profile of the expansion program compared to a company building speculatively.Second, IREN stated demand from hyperscalers, enterprises, AI developers, and frontier labs continues to exceed available and planned capacity, and the company is engaged with customers across its entire 2026 and 2027 expansion program. That is a sold-out business, not a business chasing demand.The IREN and Nvidia partnership that set the foundationThe $2.8 billion contract announcement is not the only one. It builds on a strategic partnership with Nvidia on May 7, which included a $3.4 billion AI cloud contract with Nvidia for a five-year Blackwell GPU deployment and a 5-gigawatt strategic partnership targeting IREN’s global data center pipeline.As part of that partnership, Nvidia received a five-year right to purchase up to 30 million IREN shares at $70 per share, a potential $2.1 billion investment representing meaningful skin in the game from the world’s most important AI chipmaker.Related: Jensen Huang’s staggering $4 trillion bet shifts NvidiaJensen Huang described the rationale directly in the May announcement. “Deploying these systems at scale requires deep integration across the full stack — compute, networking, software, power and operations,” Huang said. IREN brings the scale and infrastructure expertise to help accelerate the buildout of next-generation AI infrastructure globally.Co-CEO Daniel Roberts noted in the July 20 release that IREN has expanded from approximately 3 megawatts of self-built AI cloud capacity a year ago to 480 megawatts being delivered in 2026, with 1.2 gigawatts targeted for 2027. That scaling pace is extraordinary by any infrastructure standard.

Approximately 85% of IREN’s $4 billion-plus ARR target is now under contract.Michael Nagle/Bloomberg via Getty Images

IREN’s financial foundation and what Aug. 27 earnings need to showIREN’s most recent Q3 FY26 results, reported on May 7, reflected the company still mid-transition from Bitcoin mining to AI cloud, according to the company’s business update. Total revenue was $144.8 million, down from Q2’s $184.7 million, primarily driven by lower Bitcoin prices and the decommissioning of mining hardware ahead of GPU installation. Net loss was $247.8 million, largely driven by non-cash impairments of $140.4 million related to that hardware decommissioning.As of June 30, IREN held approximately $7.6 billion in cash and equivalents, the July 20 announcement confirmed. That liquidity position, combined with the customer prepayment structure on new contracts, gives the company the financial runway to execute its 2026 and 2027 capacity buildout without being dependent on capital markets at unfavorable terms.The next earnings date is estimated for Aug. 27. What investors will want to see is the first quarter where AI cloud revenue becomes the clearly dominant income driver, reflecting the contracted capacity that has been coming online throughout 2026. Given that 85% of the $4 billion-plus ARR target is now under contract, and 480 megawatts of capacity is being delivered this year, that transition should begin showing up materially in the August 27 results.IREN is up 126.73% over the past year compared to the S&P 500’s 18.61% gain, according to Yahoo Finance. The $2.8 billion contract announcement and the ARR target raise confirm that the company’s pivot from Bitcoin mining to AI cloud infrastructure is not a story being sold to investors. It is actually a business being built for them.Related: Iren’s AI ambitions surge in Q1 earnings

The Vampire Lestat episode 7 sets up Interview with the Vampire season 4 by examining Lestat and Louis’ failures and their love. They are the key to saving the world.

A documentary of LISA’s year since creating her company LLOUD has been documented in ;’Always Lalisa,’ directed by Sue Kim. It’s set to premiere at 2026 TIFF.

Every country that builds things eventually faces the same question about a cheaper foreign rival, and there are only two honest answers to it.You can wall the rival out and buy yourself time, or you can let the rival in and try to learn something before it eats you.Both answers cost money. Only one of them tells you where you actually stand.The United States picked the wall, and picked it hard. A 100% import duty on Chinese electric vehicles took effect in September 2024, according to the Office of the U.S. Trade Representative. A separate Commerce Department rule bars Chinese-linked vehicle software starting with model year 2027 cars and Chinese-linked connectivity hardware from model year 2030, according to the Bureau of Industry and Security.The practical result is that almost no Chinese passenger car reaches an American driveway, and almost none will.Europe went a different direction, and one country went furthest of all. A Spanish government report obtained by Bloomberg now spells out how far Madrid will go to keep its factories running, and the answer involves flying in Chinese workers to build the plants.Why Spain is betting its car industry on Chinese moneySpain is not a bystander in the auto business. It is the second-largest vehicle producer in Europe behind Germany, and the sector accounts for roughly 10% of Spanish gross domestic product and 9% of national employment, according to Invest in Spain, the government’s foreign investment agency.That is the context most American coverage skips. When a Spanish plant goes idle, the damage is not sector news. It is a national economic event.Spain has already lived through that. Nissan walked away from Barcelona. Stellantis (STLA) and Volkswagen (VWAGY) have spent years managing underused European capacity while demand for combustion cars falls off faster than anyone budgeted for.More Automotive:Top Toyota exec urges Japan’s automakers to uniteStruggling EV maker reels as Chapter 11 bankruptcy rumors swirlFord’s SUV profits are fueling its futureThe competition arrived anyway. Chinese brands took roughly 6% of European Union car registrations between January and April 2026, up from 3.2% from a year earlier, according to Euronews, which built the figure from registration data published by the European Automobile Manufacturers’ Association.I ran that against the same association’s May 2026 release, which showed battery-electric cars reaching 20% of the EU market, and the pattern is not subtle. Chinese share is growing fastest inside the exact segment Europe to which has legally committed itself.Madrid drew the obvious conclusion. If Chinese carmakers will sell in Europe regardless, it’s better to have them building in Zaragoza than shipping in from Shenzhen. Stellantis reached that conclusion first, expanding its Leapmotor partnership into Spanish plants earlier this year, TheStreet highlighted.

Spain is courting Chinese carmakers with plants and visas the U.S. has effectively banned.Bloomberg / Getty Images

What the report actually says about Chinese laborThe document, reviewed by Bloomberg ahead of publication, confirms that the country’s largest Chinese industrial investment will lean on workers brought in from China. The 4.1 billion euro battery plant jointly owned by CATL (HK:3750) and Stellantis will rely on “expatriate workers” through the fourth quarter of 2028.That is the part Washington would never sign. Not the factory; the visas.The report frames three joint ventures as investment done correctly, and it attaches numbers to them.The Stellantis and CATL battery plant carries a 4.1 billion euro price tag, about $4.7 billion, and more than 4,000 direct jobs.Chery’s venture with local firm Ebro Motors accounts for roughly 1,600 positions.BAIC’s tie-up with Santana Motors adds about 210.
Source: Bloomberg
The report is far less specific about the two things Spain says actually matter. On supplier localization and on transferring underlying technology to Spanish ownership, it describes a gradual process without committing to a timeline, according to Investment Monitor.That is the tell. Spain has the jobs in writing. It does not yet have the knowledge in writing.What America’s closed door is quietly costing DetroitHere is where my analysis parts ways with the usual take on tariffs.The wall works. Chinese cars are not reaching American dealerships in volume, and the people who wrote that policy got what they wanted.But a tariff protects a market, not a company. Ford (F), General Motors (GM), and Tesla (TSLA) do not only sell cars in the United States. They sell into Europe, South America, the Middle East, and Asia, where Chinese vehicles compete on price with no wall at all.Ford chief executive Jim Farley has been the loudest voice on this. He has described Chinese export capacity as a “wild card” for established automakers worldwide, according to Automotive World.Protection at home does nothing for a company in Madrid, Bogota, or Bangkok. It also does nothing to close the cost gap, because you cannot learn from a competitor you have made illegal to import.Spain will learn things over the next four years that Detroit has to buy, license, or reverse-engineer later. That is the trade Madrid just made, and it made it with open eyes.What investors should watch as Spain’s bet plays outThree markers will tell you whether Spain got value or just got used.Watch whether Spanish suppliers move from assembling imported kits to manufacturing real content. Watch whether that technology transfer language ever acquires a date. And watch whether Brussels follows Madrid or overrules it, since the European Commission has tightened scrutiny of Chinese investment, even as Spain argues for engagement.For a reader with an index fund, this is not foreign policy trivia. Your retirement account almost certainly owns Ford, General Motors, and Tesla, and every one of them competes in markets where the American wall does not exist.The question worth sitting with is not whether Chinese cars reach your local dealership. They probably will not.It is whether the companies that built the car in your driveway can still win the rest of the world, the part that never built a wall. Spain has placed its bet. Washington placed a different one, and only one of them looks smart in 2030.Related: China is becoming the auto industry’s innovation lab

AI chatbots can generate polished responses within seconds, allowing users to refine rough prompts with minimal time and effort. That speed has made AI tools a daily habit for a growing share of American workers, many of whom treat the trade-off as obviously worth it.For summarizing meeting notes, automating data entry, and drafting routine documents, AI tools offer a genuine competitive advantage in the workplace. A growing body of research, however, warns that the advantage carries a hidden cost that compounds over time without obvious warning.The cost does not appear in the tasks AI handles well, but rather in the judgment, original reasoning, and problem-solving skills it gradually displaces. Wharton’s cognitive surrender study reveals how AI overrides worker judgmentIn a working paper posted in January, Wharton marketing professor Gideon Nave and postdoctoral researcher Steven D. Shaw introduced the concept of “cognitive surrender.”They describe it as the habit of accepting AI-generated answers with little scrutiny, bypassing both intuition and deliberate reasoning.The study used nearly 10,000 individual reasoning trials in which some participants could consult a chatbot secretly programmed to deliver either correct or confidently wrong answers. When the chatbot was right, participants’ accuracy jumped 25 percentage points above the baseline set by people who worked alone.When it was wrong, accuracy fell 15 points below that same baseline, meaning AI access made them worse off than having no help at all. Participants accepted the chatbot’s incorrect answers 80% of the time, and their confidence rose 11.7% regardless of accuracy.”My biggest takeaway is that when AI is wrong, we see that people end up performing worse than if they had no AI at all and become more confident in their wrong answers,” Nave told Forbes.Early-career workers face the steepest risk from AI-driven skill erosionThe Wharton data raises a pointed concern for workers in the early stages of their careers, when judgment and expertise are still developing. Dorothy Leidner, a business and ethics professor at the University of Virginia, has warned that the risk hits hardest for those who lack foundational experience.”The risk is higher for younger workers who do not have a deep expertise to help them use AI as a thought enhancer rather than a thought generator,” Leidner told CNBC.More AI:Goldman Sachs has a blunt message for AI stock investorsMicrosoft CEO sends a blunt warning on AI and the tech ecosystemThe next AI infrastructure race has nothing to do with chipsMohammad Hossein Jarrahi, an information science professor at the University of North Carolina at Chapel Hill, uses a related concept he calls “cognitive debt.” The term describes the capability that erodes each time an employee lets AI make a decision that requires independent judgment, Jarrahi explained to Human Resources Director (HRD).He warns that the most dangerous category of AI delegation involves the hard judgment calls where independent thinking develops through repetition and experience.

Early-career workers face greater AI risks as overreliance weakens critical thinking, judgment, and the expertise built through experience over time.Morsa Images/Getty Images

Researchers warn that AI overuse erodes both originality and brain functionThe concern extends beyond individual skill loss into the quality and originality of the ideas employees produce for their organizations. Sandra Matz, a professor of business at Columbia Business School, has published research showing that AI-assisted choices become increasingly predictable and less varied over time.The underlying ideas converge toward the same outputs, even when the surface-level language still appears diverse, the study found.”When employees use it to outsource critical thinking or creative reasoning, there’s a real risk of gradually falling behind and becoming complacent,” Matz told CNBC.Adam Green, a neuroscience professor at Georgetown University, reinforces the concern from a biological perspective through his laboratory research. Unpracticed cognitive skills undergo measurable physical deterioration in the brain through the same neuroplasticity that builds them, Green explained to HRD.Where the line falls between helpful AI and a career liabilityThe key takeaway is not to stop using artificial intelligence but to use it more thoughtfully and carefully. Americus Reed II, a marketing professor at the Wharton School, draws a clear boundary between productive and destructive uses of AI in comments shared with CNBC.It should remove friction, not replace judgment. The goal is to spend less time processing information and more time creating meaning.Reed points to summarizing information, analyzing data, and organizing ideas as areas where AI adds clear value, while flagging judgment-intensive work as the terrain where users need to keep their reasoning in the loop.The 2026 International AI Safety Report, backed by expert nominees from more than 30 countries and international organizations, found that reliance on AI tools can weaken critical thinking skills and encourage “automation bias.”The pay premium on AI skills is rising, but AI fluency alone isn’t the whole storyFor workers navigating a job market absorbing AI at an accelerating pace, the central takeaway from this body of research is direct. The researchers cited above converge on a shared warning: The tools themselves are not the threat, but uncritical dependence on them can quietly erode the judgment and reasoning skills that make workers hard to replace.Job postings that list AI-related skills already command salaries roughly 28% higher than those without them, research from labor analytics firm Lightcast has indicated. The workers who benefit most from AI in the long run will be those who combine AI skills with their own critical thinking and judgment.The Wharton study’s closing recommendation is therefore blunt. AI systems need to prompt people to think, not to think for them, Shaw and Nave concluded. Until the tools are redesigned to do that, Shaw and Nave argue, the responsibility for maintaining independent judgment rests with the individual user.Related: J.P. Morgan examines if AI is really after your paycheck

So far, the 2026 housing market hasn’t been all it was originally cracked up to be.I’ve been covering the housing market for years, and although I’m disappointed that fewer people are buying homes right now, I’m also not at all surprised.Stubbornly high mortgage rates. Home prices that are unrealistic for many buyers. A volatile geopolitical climate that affects the market in numerous ways.These are all factors contributing to falling pending home sales, according to a Redfin report. In the four-week period ending July 12, pending sales in the United States dropped by 2.2%.This marked the first pending home sales decline in a month.”Data on home sales provides indicators about current housing demand,” Daryl Fairweather, chief economist at Redfin, told TheStreet. “When fewer homes are selling, buyers have more negotiating power, while indicating to sellers that they may need to lower their price or offer concessions to sell their home quickly.”Fewer pending sales impact both homebuyers and sellers, just in different ways. And if the trends continue, there could be a longer-lasting affect on the real estate market.Redfin clarifies that America is still a buyer’s marketAmerica is still a buyer’s market, according to Redfin, meaning there are more homes for sale than people looking to buy.But new listings are decreasing, too.Week-over-week new home listings fell by 1.2%. This landed new listings at their lowest point since the beginning of 2026.Sellers who don’t have to move are choosing to stay where they are due to low buyer demand. It doesn’t benefit them to list their houses only for it to sit on the market for a long time, which could result in a price cut or less negotiating power on their end.Related: Fannie Mae predicts shift in mortgage rates, housing marketDuring the four-week period ending on July 12, there were 350,510 new listings. There were also 3.4 months of supply, which is down 0.2% year over year.Months of supply refers to the amount of inventory divided by the rate of home sales. Redfin states a “balanced” number is four to five months of supply. A lower number typically indicates that it’s a seller’s market.However, I’ll be the first to admit that the real estate market doesn’t always follow the “rules.” The low 3.4 months of supply might make it look like America is a seller’s market on paper, but in reality, we are more of a buyer’s market.It’s just that fewer pending home sales hurts multiple facets of the housing market right now. Including inventory.

Fewer pending home sales impact both buyers and sellers.Grace Cary / Getty Images

Fewer pending homes sales should help buyersJohn Walkup, co-founder at UrbanDigs, homed in on the importance of pending sales versus closed sales. He explained that closed sales data is based on decisions buyers made months ago.But pending sales data reveals more recent home-buying activity and is a better representation of how the housing market is performing currently.That’s why Redfin’s updated numbers on pending sales data is so important. It helps us understand that current activity is slowing, even if we see other headlines that overall sales are up.”So even if the ‘sales are still up this year’ narrative holds, the drop in pending deals suggests demand may be losing momentum below the surface,” Walkup told TheStreet. “For buyers, that means less competition and possibly more leverage. For sellers, it means days on market and no room for aspirational pricing.”More Housing Market:Zillow data shift sparks major controversy in housing marketGoldman Sachs issues major prediction for U.S. housing marketRedfin reveals change in housing market, home sales”Lower home sales can be good for buyers if it’s a result of lower demand, especially in the short term,” Michael Weiner, a real estate agent at Coldwell Banker Warburg, told TheStreet. “Basic macroeconomics should result in lower prices from those who need to sell, especially in a local market.”Homebuyers are in a sticky spot. On one hand, the main reason pending home sales are down is that affordability is so bad. High costs put homeownership out of reach for many Americans.But if you can still afford a home, then lower pending home sales puts you in a better position and can save you money. When there’s less competition, you have more power to negotiate with the seller and face less competition with other buyers.Low pending sales can be a win for homebuyers — but not all homebuyers can win in this situation.What happens if long-term pending home sales stay low?Fewer pending sales makes housing more affordable for buyers. So, if home sale numbers are low for a long time, prices should also stay low, right? Maybe home prices will even decrease.Well, not necessarily.”Widespread declines typically leave homebuilders to scale back capacity, leading to lower supply,” Weiner told TheStreet. “Lower supply also leads to more people staying in their current homes, even if they want to move, making it more expensive to buy the few homes available.” “It can lead to a vicious circle in which entire markets could become nearly frozen, as most people stay in place if they can,” he said.In a homebuyer’s perfect world, pending home sales wouldn’t spike — that would signal more competition, which could lead to bidding wars and higher sales prices. But neither would they continue to decline for months and months. That would make it difficult to find a home at all.Buyers and sellers alike should keep an eye on upcoming pending home sales data. Redfin updates its housing market data every Thursday, and the company will post the next monthly update on August 10.For information on your current housing market, talk to a local real estate agent. Pending home sales in your area may be higher or lower than the national data shows.Knowledge about ongoing pending home sales trends can help homebuyers know what to expect as they begin house hunting. And it can help sellers decide whether to list their homes soon.Related: Why mortgage rates are spiking again and what to do

There is a rule about money most of us learn early. You pay for the thing, not for the room the thing happened in.A gallon of milk costs what it costs whether you buy it at a corner store or a warehouse club. If the corner store charges triple, you notice, you complain, and eventually you stop going.Health care in this country runs on the opposite principle. The same X-ray, taken on the same machine and read by the same radiologist, can carry two very different prices depending on which corporate entity owns the sign above the door.That gap did not appear by accident. It was written into the payment rules, and hospitals responded the way any business responds to a pricing loophole. They bought the independent practices, hung a new sign, and started billing at the higher rate.Nobody voted for that. It happened one acquisition at a time, and it ended up inside your premium and your coinsurance.Now Medicare wants to close part of the gap. The Centers for Medicare and Medicaid Services, or CMS, released its proposed calendar year 2027 outpatient payment rule on July 2, and hospitals responded within hours.Why the same X-ray costs more inside a hospitalMedicare pays for outpatient care through two systems that do not agree with each other. Services delivered in a Hospital Outpatient Department, known as an HOPD, are paid under one fee schedule. Identical services in an independent physician’s office are paid under another, and the hospital version pays more.That differential raises costs for the government and for beneficiaries and “encourages excessive hospital consolidation,” according to the Committee for a Responsible Federal Budget.Related: Dave Ramsey, Fidelity deliver key alert on Medicare, 401(k), IRAThe fix has a name that has floated around Washington for close to a decade. Site-neutral payment means Medicare pays the same amount for the same service regardless of where it happens. Applied across all of Medicare, it could save the federal government almost $200 billion over ten years, per the same budget watchdog.I have read a lot of Medicare proposals, and this is the rare one where the underlying logic is not seriously contested. An X-ray does not become a better X-ray because it was taken in a building with a cafeteria.What has always blocked reform is that hospital revenue is a real thing in real congressional districts, and there is no painless way to pull several billion dollars out of it. What the 340B drug payment cut would do to hospital revenueThe bigger number in this rule is not imaging. It sits inside a drug discount program created in 1992 called 340B, which requires manufacturers to sell drugs at steep discounts to hospitals serving low-income communities.Here is the mechanic that matters: Hospitals buy those drugs at a deep discount, then bill Medicare at the Average Sales Price, or ASP, plus 6%. The spread between what they pay and what they collect has become one of the most dependable revenue lines in American health care.CMS wants to end it. The proposal would pay for 340B drugs at ASP minus 33.4%, starting in 2027. This would cut Original Medicare drug payments by $4.55 billion and beneficiary drug payments by $1.15 billion in the first year, according to CMS.More Medicare/Medicaid:Medicare’s costliest gap threatens retirement savingsMedicaid’s 5-year rule catches families off guardMedicare Advantage payments rise. What changes nextHospitals are not treating this as a technical adjustment. The cut “will make drugs less affordable for America’s most vulnerable patients,” said Ashley Thompson, senior vice president of public policy analysis and development, per the American Hospital Association.A second blow lies underneath. CMS also proposes to accelerate its clawback of earlier 340B overpayments by cutting the outpatient conversion factor 3% a year instead of 0.5%, finishing recovery by 2029 rather than 2041.The rule “delivers multiple blows to hospitals already operating under immense financial strain,” Premier advocacy vice president John Knapp said in a statement, as reported by HFMA.Here is how the savings math breaks down across the two site-neutral policies now in play:Site-neutral payment for outpatient imaging would save Medicare roughly $7 billion over ten years, according to CMS estimates cited by the Committee for a Responsible Federal Budget.That same policy would save closer to $10 billion while cutting beneficiary premiums and cost sharing by another $7 billion, based on scoring from the Congressional Budget Office.The 2026 final rule’s site-neutral policy for drug administration saves about $7 billion federally and $5 billion for beneficiaries over a decade, per the Congressional Budget Office.Together, the two policies would cut national health spending by almost $20 billion over ten years, according to the Committee for a Responsible Federal Budget.

Proposed 2027 Medicare rule cuts 340B drug pay 40% and expands site-neutral imaging.FS Productions / Getty Images

What the Medicare outpatient proposal means for your costsThis is where my analysis turns less cheerful than the headline numbers suggest. Statute requires most of these changes to be budget neutral, so the roughly $4.85 billion pulled out of 340B drug payments does not reach the Treasury or the Medicare trust fund. It goes back to hospitals as an 8.44% bump in payments for non-drug outpatient services.Savings get recycled into Part B payments instead of shoring up a program facing rapid cost growth, the Committee for a Responsible Federal Budget noted, which is why the group wants Congress to codify the change rather than leave it with regulators.So what actually reaches you? Two things, and neither is small.Your Part B coinsurance is 20% of whatever Medicare allows. When the allowed amount on an infused drug drops by nearly 40%, your share drops with it. That is the $1.15 billion, and it lands hardest on people receiving chemotherapy and biologic infusions.The imaging change works the same way for anyone who has had a scan at a hospital-owned clinic, and it feeds into premiums, the quiet lever that decides how much of your check the government takes back each January.None of this is settled. Comments close Aug. 31, a final rule is expected later this year, and the last time CMS attempted a 340B cut like this, hospitals sued and won at the Supreme Court in 2022. This version rests on a survey of actual hospital acquisition costs, which is precisely the evidentiary gap the justices flagged, according to the Federal Register notice.Watch two things between now and January. First, whether CMS carves out children’s hospitals, cancer hospitals, and rural sole community hospitals, all of which it has flagged for comment. Second, whether the roughly 618 procedures being added to the ambulatory surgery center list start pulling volume off hospital campuses, which would matter for investor-owned operators including Tenet Healthcare (THC) and HCA Healthcare (HCA), according to Holland & Knight.The site-neutral argument has been running since 2015 without resolution. This is the first version that arrives with a survey attached, and that changes who has to prove what.Related: Can Medicare Help You Get GLP-1s? Requirements You Need to Know

Old Petty has officially opened at Cabot Highlands resort in the Scottish Highlands.

Murder haunts King’s Landing as the armies of Team Green and Team Black march toward war in the latest episode of House of the Dragon on HBO.

Enterprise AI is facing an ROI paradox. While throwing more compute at the strongest foundation model works well in product experiments, the costs become unbearable when the product is deployed in production.A new paper from researchers at Writer provides a solution that is accessible to engineering teams. The study takes a systematic look at optimizing the different components of the orchestration layer that wraps around the foundation model, aka the AI harness. By optimizing the harness, the researchers show dramatic reductions in tokens per task, a drop in cost-per-successful-task by up to 61%, and quality that holds steady, all without changing the underlying foundation model.Because the harness is fully under the developer’s control and requires no model fine-tuning, engineering teams can apply these findings to build highly cost-efficient AI applications.The ROI crisis of tokenmaxxingThe current state of AI engineering is plagued by “tokenmaxxing,” an industry trend where developers rely on massive context windows and brute-force token consumption as a substitute for good system design. Rather than engineering elegant workflows, developers have imported a reflex from traditional software development: generate, run, fail, stuff the error and more context back into the window, and retry. “Teams tokenmaxx because it’s the cheapest fix in the moment, and because it’s literally how most engineers work today,” Waseem AlShikh, CTO and co-founder of Writer, told VentureBeat. Because this approach succeeds often enough on coding tasks, it has become the default reflex for every other agentic workload. The danger is that per-token price drops mask the underlying inefficiency. “Your invoice is tokens-per-task times price-per-token, and most teams only watch the second number,” AlShikh said. “In agentic workloads, tokens-per-task compounds — every loop iteration re-transmits the growing context — and it compounds faster than prices fall. The price cut becomes an anesthetic. It masks the fact that the loop itself is bleeding.”Tokenmaxxing leads to several enterprise failure modes. Teams route simple tasks to premium frontier models by default. They use the LLM as a lazy search index, stuffing the context window with raw documents instead of retrieving exact answers. Most destructively, they build unconstrained agentic loops that spiral out of control when the model encounters an error. Because output tokens cost significantly more than input tokens across all major model providers, inefficient task execution acts as a silent budget killer.The industry has introduced several efficiency techniques to curb these costs, but they largely fall short because they treat the model in isolation: Prompt compression condenses input text to save space, but ignores how the system sequences those inputs across complex workflows. Budgeted reasoning caps the computational steps a model can take, which often degrades output quality if the workflow isn’t intelligently routed. Terse coding forces models to output minimal code to save output tokens, but does nothing to solve inefficient tool calling. Speculative decoding uses a smaller draft model to speed up a larger model’s text generation, optimizing inference speed while failing to address bloated agent architectures.These efforts fail because they optimize the engine while ignoring the transmission. They do not look at the orchestration layer, leaving underlying architectural inefficiencies unresolved.Unpacking the harness: the levers of efficiencyThe harness is the orchestration layer that routes, formats, and turns the underlying LLM into a working system.The core levers of harness optimization include system prompt caching, interaction history compaction, tool management, retrieval strategies, and error management. These are the most accessible intervention points for engineering teams looking to improve AI performance. As the Writer researchers note in the study: “If the harness is the layer that composes model calls into work, it is also the layer that sets the price of work.”Historically, developers have treated the harness as disposable glue code designed simply to connect an API to a user interface. The study signals that the harness must now be treated as a first-class object: a primary software artifact that requires its own testing, versioning, and rigorous design. For enterprises, this reframes the “own-versus-rent” decision. “Enterprises spend months on model evaluations and then rent their orchestration off the shelf — which means they’re optimizing the smaller lever and outsourcing the bigger one,” AlShikh said. “Whoever owns the harness owns your unit economics, and an open framework tuned for demos is not tuned for your invoice.” Inside the experimentsTo isolate the impact of the orchestration layer, the researchers ran experiments on six foundation models spanning multiple vendors and weight classes: Claude Sonnet 4.6, Gemini 3.1, Gemini Flash 3.5, Qwen 3.6, GLM 5.1, and Writer’s own model, Palmyra X6. Their experiments compared a frozen, conventional production agent loop against the finished Writer Agent Harness on the same 22 locked enterprise tasks, spanning capabilities like grounding and retrieval, multi-step workflows, tool use, and content generation. By holding the models and tasks constant, they could isolate the effects of the orchestration layer itself.The optimized harness drove a significant drop in costs, cutting the blended cost per task by 41%, from 21 cents to 12 cents. This was largely achieved by slashing token consumption, with the number of tokens per task falling 38%, from 14.2k to 8.8k.The harness is designed to delegate tasks like search to specialized sub-agents. A sub-agent receives only the tool and the specific query it needs, retrieves the exact data, and returns a capped, clean summary to the main agent — keeping the primary context window from filling up with raw search results.Task success rates held steady even as token use fell — moving from 78% to 81%, a gain the researchers describe as directional rather than statistically significant at their sample size, meaning quality didn’t suffer even as costs dropped.End-to-end task latency also dropped significantly, reducing the median wall-clock time by 44%, from 48 seconds to 27 seconds, due to prompt caching and the elimination of dead-end reasoning loops.However, the researchers also found limits to multi-agent orchestration. Smaller models like Gemini Flash 3.5 and Qwen 3.6 scored well below a usable reliability threshold on sub-agent delegation tasks (0.45 and 0.42, respectively) — the capability simply isn’t dependable yet on lighter-weight models.Sub-agent orchestration only crossed a usable reliability threshold on the two strongest models tested: Writer’s own Palmyra X6 (0.86) and Claude Sonnet 4.6 (0.85).The developer’s playbook: actionable takeaways and tradeoffsThe findings from the study translate into a playbook for enterprise developers building agentic workflows at scale. The first step is to implement what AlShikh calls the “Two-Zone Prompt” and “Context Offloading.”Structure for system prompt caching (The Two-Zone Prompt): Modern LLM APIs offer prompt caching, but developers must structure their payloads correctly to trigger it. Developers must separate the “stable zone” from the “volatile zone.” Place static, unchanging elements (e.g., core rules, large tool schemas, and standard operating procedures) at the top of the prompt. Dynamic elements, such as the specific user query or recent conversational task state, must be appended at the bottom. This ordering allows the harness to reuse the cached prefix across hundreds of calls. “That single separation makes prompt caching actually work and stops you from re-paying for the same instructions on every one of an agent’s thirty steps,” AlShikh said.Manage context with Context Offloading: Avoid context stuffing, where every turn of a loop is appended into a monolithic prompt until the window maxes out. Instead, move history and intermediate artifacts out of the window into retrievable storage, and pull back only what the current step needs. If possible, delegate tasks to single-purpose sub-agents to avoid context bloat. As AlShikh points out, “the biggest line item in agent spend isn’t reasoning — it’s re-sending things the model has already seen.”Build resilient loops and redefine KPIs: Unmanaged agent loops drain API budgets rapidly. Teams must begin tracking Completions Per Million tokens (CPM) to understand their true task costs, but the harness itself must contain physical guardrails. “The core principle is that you never ask the model to police its own spending,” AlShikh said. “The fence has to live below the model, in code, on your side of the API.” This requires three hard checks:Hard per-task token budgets: The run terminates when the budget is spent, no exceptions.Generation fencing: Caps on steps, tool calls, and recursion depth to stop non-converging agents. Failure-spend governance: Cap what a run can spend after its first failed validation so a failing task doesn’t become your most expensive task.Avoid unnecessary complexity: Optimizing the orchestration layer comes with engineering overhead. If you’re in the prototyping and exploration stage, that overhead isn’t justified — iterate fast with a strong model and a light harness. Once you’re scaling to millions of requests a day, the savings from harness optimization become substantial.However, teams must be aware of “harness leverage.” Adding structural scaffolding requires the model to hold and obey that context. If a model is too small, it will spend its limited capacity parsing the scaffolding instead of doing the task, causing accuracy to drop and tokens to rise. The rule for adding complex orchestration features is strictly mathematical: “If a feature adds more coordination tokens than it removes task tokens for that specific model, cut it,” AlShikh said. “Nothing in the harness is free.”The future of the enterprise harnessThe era of tokenmaxxing and treating context windows like bottomless buckets is coming to an end. Throwing more compute at poorly designed systems is not a viable strategy for companies that need to demonstrate a return on their AI investments. As foundation models evolve to absorb planning, tool selection, and multi-step reasoning natively into their weights, the role of the harness will shift from compensating for model weakness to enforcing enterprise policy.”What never moves into the model is the ‘allowed’: budgets, permissions, data boundaries, audit trails, deterministic kill-switches,” AlShikh said. “Five years from now, the harness will be thinner but more important. There will be less scaffolding and more governance. However capable the model gets, someone external to it still has to define what it may spend, see, and touch. That layer belongs to the enterprise, and it should never be rented.”