Costco is soon to offer Medicare plans as part of an innovative, new health insurance venture.
SCAN Health Plan is one of America’s largest Medicare Advantage nonprofit plans that focuses on seniors health care. The insurance company made the announcement Tuesday.
In the next few years, both companies plan to launch “a suite of insurance products aimed at seniors”.
The suite, which is subject to regulatory approval, could include an improved pharmacy experience, Medflex benefits for over-the counter pharmacy, coverage of vision care, and audiology, hearing, or other benefits.
COSTCO ADDS HOT FAN FAVORITE FOOD COURT MENUS AS SHOPPERS DELIBERATE TASTE & VALUE
The Wall Street Journal reports that the plans will also be available via insurance agents, websites and Costco.
Companies haven’t revealed the exact date when plans will be available. Regulatory review and approval is still in progress. SCAN serves 33 counties across California, Arizona Nevada, Texas New Mexico Washington.
SCAN Health stated that the initiative is designed to “address the disjointed experiences many seniors have accessing services and care” and “will bring more value and a better experience for seniors as they navigate through their health insurance”.
COSTCO MAKES A PAYMENT MODIFICATION THAT COULD SPEED THE CHECKOUT PROCESS FOR MEMBERS
In a press release, Dr. Sachin JAIN, CEO of SCAN Group, and SCAN Health Plan said that the future success of Medicare Advantage depends on partnerships which make healthcare easier to access, better connected, and seamlessly integrated with everyday life.
The older adults we serve want healthcare which is more accessible, responsive to their requirements and grounded in trusted organizations. We will use our expanded partnership with Costco to find new and innovative ways to keep people healthy, independent, while providing the same quality, value and service that both organizations are well known for.
COST COSTCO WHOLESALE CORP. 961.35 +7.85 +0.82%
Costco’s CEO Ron Vachris said that the partnership between Costco and SCAN is an extension of Costco’s commitment in providing value for its customers.
Costco’s commitment to its members has been consistent for more than forty years. They have earned the trust of our members by providing them with quality essential goods at a fair price and expanding their health services.
CLICK HERE TO GET FOX BUSINESS ON THE GO
SCAN is a partner that we have chosen to help us deliver better care for our seniors. This decision represents an extension of this commitment. Our shared understanding has helped us to better serve the senior population. “Deepening our partnership allows for us to create value and build upon that foundation.”
FOX Business contacted Scan Health to get more information.

California Attorney General Rob Bonta stated Tuesday, following the first day’s arguments, that the case is primarily focused on “civil punishments, restitution, and distortion” and not money.
Bonta told a post-session press conference that “this is not a case of damages.”
Trial at Oakland’s federal courthouse is anticipated to last between six and seven weeks. Bonta, along with attorneys generals from Colorado, New Jersey, and Kentucky, are leading the Meta trial. They represent a group of bipartisan 29 AGs that allege the features and design of Facebook and Instagram encourage addictive behaviors in children, which are harmful to their mental health.
Meta’s attorneys have warned before that the combined case may result in penalties of up to $1.4 trillion or about equal to Meta’s current market capitalization. Last week, state AGs informed presiding federal judge Yvonne Gonzalez Rogers that 200 billion dollars is a better estimate.
Gonzalez Rogers, however, will decide the fate of Meta. The same set-up was used for Elon Musk’s trial against Sam Altman in early this year.
Lawyers for the states claim that Meta has violated state and federal laws, including the Children’s Online Privacy Protection Act (COPPA) and other consumer protection laws. Bonta stated that the lawyers of the states “submitted documents to the court,” which included statistics about the number young people who use apps such as Facebook and Instagram to help determine civil penalties.
Bonta stated on Tuesday that “we are not asking $1.4 trillion,” he said. Bonta said that it was the “highest number possible that could be calculated using this data” and “it’s a Meta estimate.” Meta denies this claim.
Bonta stated that the plaintiffs will “ask for an significant financial gain as a result of the total civil penalties in this case, which is an important part and appropriate of the lawsuit.” The judge has discretion to decide what this will be.
This trial is a result of a 2023 lawsuit. Megan O’Neill, California’s Deputy Attorney-General laid out Meta’s claims during opening arguments. O’Neill’s arguments were similar to those of other social media cases, such as the Los Angeles personal injuries proceedings Meta and Google’s YouTube both lost in March. She focused on the allegations that executives, like Meta CEO Mark Zuckerberg, made public statements that contradicted internal company research about the mental health of children, and misrepresented their true extent of harm.
O’Neill, a member of the advisory jury, said: “Meta is one of the most powerful and largest companies in the World. It has conducted a deceptive campaign, misleading its users, legislators, educators, parents, etc.”
Paul Schmidt, a Covington & Burling attorney who represents Meta, contested O’Neill’s arguments and disputed that Meta executives had misled the public by their statements regarding child safety.
As you listen to the evidence, we’ll ask you: Does this support the claims they make? Schmidt told the jury.
The state attorneys want Meta, in response to the allegations, to delete the personal information of children younger than 13 years old, as well as the algorithms and models that were trained using this data, and remove design features like autoplay, infinite scrolling, and beauty functions.
This trial comes after a recent victory by New Mexico AG Raul Torrez against Meta, which resulted with the social media company being ordered to pay $567,000,000 into an abatement account as part of the 2nd phase of a lawsuit involving allegations of child sexual abuse. In March, a jury had already decided that Meta must pay $375 in civil penalties to New Mexico for its violation of the unfair practices act.
Bonta pointed out that Meta was not the only company targeted by state attorneys general. Bonta said that several states have “already pending lawsuits against TikTok” and he is “also worried about YouTube as Snap.” He said that Meta was “just the first in line” in an effort to find “industry-wide solutions” in all of California.
Who goes first? “Who goes last?” Bonta stated that “ideally, everyone would go together.” It’s impossible. They would commit to reforms that will keep all children safe.
Watch the Landmark Meta Trial. What you need to know

Hot Chips will have more information on CS-4. CS-4, their 4th-generation rack is built on the same 3rd generation 5nm Wafer-scale engine WSE-3. CS-4 is twice as fast as CS-3, thanks to increased clock frequencies and power consumption per wafer.
All of this translates to CS-4 having the ability to double tokens/s/user from CS-3 at a cost that is similar. Customers will be able to double their token revenue for the same amount of hardware. Not only are tokens faster but also the time it takes to get to market. The rack architecture has been redesigned and is more modular. This allows for better manufacturing capabilities and deployment. A new I/O Module will allow for open heterogeneous inference architectures to be developed. By pairing the CS-4 with HBM systems, these disaggregated setups can help to overcome the memory limitations of the CS-4.
The same wafer with double the clock
Cerebras has doubled the clock speed of the CS-4 to extract twice the performance. The CS-4 has improved cooling and power technology, allowing it to deliver more power onto the wafer. Cerebras, while staying with the same silicon (5nm) may not sound impressive, can double the most important metric: memory bandwidth. This doubling should result in a near-doubling of tokens/sec/user, all other things being equal. And this is exactly what Cerebras’ customers are looking for. The clock speed doubling will also double peak FLOPs theoretically and WSE parallel I/O. This allows the CS-4 upgrade from 1.2Tb/s off-wafer to 2.4Tb/s. The 44GB SRAM per wafer remains unchanged, since it is determined by how many SRAM bits are available per wafer. We’ll need to wait until the next-generation silicon is available before we see improvements. The main disadvantage of using the WSE-3 is that the memory is limited per wafer. This is a key feature of Cerebras’s architecture.
The wafer, as we have described in the previous article about Cerebras has a unique architecture because it uses SRAM. This makes the wafer well-suited to run kernels that require low Arithmetic intensity, like low batch size decoding.
Improve your network with Network Enhancements
The Wafer interface is an improved FPGA card used to convert Cerebras proprietary ethernet I/O into standard ethernet. On the below image, you can clearly see two I/O modules that come off of the north and the south side of the wafer.
Cerebras is able to upgrade this I/O module on the field, allowing it to adapt quickly and easily for new standards in networking without having to redesign its chassis. It may seem like a minor change, but it has a lot of implications. The CS-4 can now interface more easily with other systems to create disaggregated setups. In the same manner that Nvidia positions the Groq LPUs, pairing the CS-4 and HBM-based XPUs to create a network with disaggregated feed-forward attention is one solution for the CS-4’s limited memory. Below, we will write about it in more detail. It seems that this is designed specifically for AWS, which would like its EFA NICs to interact with Trainium servers to disaggregate inference.
A new pipeline for low-latency processing of packages reduces latency in the two layer fat tree network using Arista ethernet switch technology to 3 microseconds, down from 5 microseconds. Direct wafer-to-wafer links can also be made instead of going through the switched networks, which reduces latency further to two microseconds. Direct wafer paths can also be configured, so the FPGA is able to switch data between wafers.
We consider this a moderate improvement. This 3us bottleneck and bandwidth limitation is preventing parallelism in setups like EP or ETP, where the expert layers are spread across multiple wafers. The latency of token dispatch from the router to the expert layer is critical, as well as the problem with expert imbalance.
The Backpack rack
The biggest change at the system-level is the physical design. Cerebras splits the rack in two halves, one for power and another to compute. Each half is packaged into modular “backpacks” that plug together. A CS-4 rack can hold three backpacks, up from just two wafers on a CS-3 rack. Data centers are now built with racks that are fully liquid-cooled, so the cooling infrastructure like the pumps and heat exchangers is also removed.
The backpack enclosure is a vertical enclosure which is divided into separate power, cooling and I/O modules, as well as WSE-3, making the entire system easier to produce than CS-3. WSE-3 is positioned vertically, with its power side facing forward. Power modules deliver power through the wafer’s front, the same side as CS-3. Cooling the wafer is done from the back side. The I/O Modules are mounted on the wafer’s top and bottom edge, forming a rectangle with the wafer.
This backpack allows for a more efficient deployment. The rack can be set up with power modules, and then the backpack is simply plugged into the rack. The upgrade from three wafer engine per rack to higher clock speeds. A single CS-4 rack can reach 125-135 kW TDP. This is around or slightly less than double the power of a CS-3. This means that performance/W is at most a marginal improvement over CS-3.
Cost should be the biggest improvement from generation to generation. The CS-4 has fewer components, which should reduce the BOM. We think that the BOM for the CS-4 could be similar to the CS-3. This means that customers will get nearly twice the amount of interactivity, and token revenue at a similar TCO.
Cerebras has chosen 43 PB/s as the number they want to use for CS-4. This is roughly 2,000x greater memory bandwidth compared to Nvidia Rubin. This number will dominate the coverage for this launch because on-wafer SRAM capacity scales with power.
Cerebras is claiming that despite 2,000x greater memory bandwidth it can still improve interactivity by up to 30x compared with GPUs. It’s calling this new performance level “ultrafast”. We think CS-4 models will reach a tok/sec/user of 4,000 on the frontier model, and CS-3 will be at 2,000. We expect Blackwell GPUs to continue reaching a theoretical maximum of 200 tok/sec/user, which no one runs on. A more realistic top speed is 100 tok/sec/user with reasonable concurrency. This looks to be 20-40x greater interactivity, why then not make it “up to 40x faster”? It seems fair.
Parallelism strategies
Cerebras has continued to use this system to perform pipeline parallel inference because a WSE is not equipped with enough SRAM to store the weights of an entire model. Each MoE expert will be interleaved on a wafer for CS-4. The default parallelism for pipelines is not the same as GPUs. Tensor parallelism and expert parallelism tend to be used to make large models fit into HBM. Cerebras’s position has been that using HBM on the GPU to store weights would be slower, power-hungry, and costlier than doing it all in one wafer.
It really does depend on the parallelism strategies you use when you compare a WSE cluster to a GPU cluster in terms of performance, cost and power consumption. The GPU’s configuration ranges are vast (from low interactivity/high throughput to high interaction/low throughput), while wafers offer a more limited range. The only configuration considered is high interactivity/low performance.
We are interested in NVIDIA TileRT’s high-interactivity/low-throughput configurations for GPU clusters. In our TileRT piece last week, we discussed some of these issues.
Cerebras doesn’t need to shred individual experts for MoE models because the wafer is so large. This applies even when they are running frontier models like GPT 5.6 sol. We expect that OpenAI, for example, will run the 5.6 Sol model at the 256k context instead of the 1M context, in order to save money on keeping the KV Cache installed on the wafer.
The cost to support long-term context inference, as we explained in our Cerebras IPO Article is enormous. The majority of people are still aware that the memory needed to perform a forward pass on a model depends on the number of parameters. It is still a secret, but the memory needed to run KV Cache increases proportionally with the number of users concurrently and their average/maximal request size (which are determined by the model’s context window). Memory is needed to run a model that has a big context window and supports many users.
We find, when we do a simple calculation of how many Cerebras WSEs are needed to run the DeepSeek Pro 1.6T model, that it takes around 20 Cerebras WSEs to run the model with 1M ctx. At a concurrency reasonable of 256, we need around 40 Cerebras WSEs. This is over $20M in CAPEX, and 1MW power usage before you get to a front pass for a frontier-model.
You can find this analysis on the public website of tokenomics:
After this initial investment, you will see some impressive aggregate performance metrics. Now let’s dive into it.
Bet on disaggregation
Cerebras positions CS-4 from the start as being designed for heterogeneous, open inference. It is currently working alongside AMD Trainium and AWS Trainium, with more partners to come. Cerebras is the decoder chip in all of these configurations because its roofline doesn’t allow for prefill compute bound. It claims to support attention feedforward (AFF) disagg as well as the traditional prefill-decode disaggregation.
The bandwidth of SRAM architectures cannot be compared to HBM-based software approaches that try to emulate dataflow execution. All heterogeneous Disagg configurations have a double edge, however, as the ratio between Prefill and Decode is set on the date that your hardware order is placed. A fleet of TPUs or GPUs can also be dynamically assigned into different ratios, as the workload profiles change over time. In the real world workloads change over time. As models thought more, decoding costs increased. Then agentic caches drove prefill costs lower while decoding costs remained unchanged. It is not likely that a single P:D ratio will be optimal over the 5-year lifespan of this system.
Read our TileRT blog to get a deeper understanding of the dynamics and how infrastructure teams design inference clusters that are high-performance.
Nexus and CS-5 Roadmap
Cerebras co-designed a next generation wafer scale engine along with a new platform called Nexus. This means that the CS-4 chassis is likely to carry over into future chip generations, without major mechanical redesign. According to the company’s roadmap, it is committed to delivering 2x more performance each year with a goal of achieving 20x improvement in throughput by 2027.
Cerebras continues to invest in redundancy at the tray level, yield harvesting, cores, channels and error recovery on the chip.

Investor’s Business Daily contains information and education only. It is not intended to be an offer, recommendation or solicitation for the purchase or sale of securities. We have obtained the information from reliable sources, but cannot guarantee its accuracy, relevance, or timeliness. This includes information in closed captioning. The past performance of investments is not a guarantee or indication of the future. The authors/presenters of the articles may be owners of stocks that they are discussing. No representations are made or warranties given regarding the suitability of any investment strategy or any securities. The information is subject to changes without prior notice. Please see our Terms and Conditions for information about our service.
*Real time prices according to Nasdaq’s Last Sale. Not all markets provide real-time quotes and/or trading prices. Data on ownership provided by LSEG, and estimated data by FactSet.
Investor’s Business Daily, Leaderboard, MarketDiem, MarketSurge, IBD, IBD Digital, IBD Live, IBD Weekly, Investor’s Business Daily, and other trademarks are owned by Investor’s Business Daily, LLC.
Investor’s Business Daily, LLC. All Rights Reserved.

American Airlines’ passengers will soon see some major changes as they take their seats. The carrier is undergoing a complete overhaul of its onboard service.
According to the company, more than 800 narrowbody planes will receive new entertainment systems with 4K screens, Bluetooth and USB-C charging.
American announced the updates in a blog post posted on X.
Retrofitting existing jets with screens is expected to start in 2028.
CHICK-FIL-A REIMAGINES SOUTHERN BREAKFAST CLASSIC AS AMERICANS PILE ON THE PROTEIN
According to American, the retrofit program will last until early 2030.
This is a significant shift for Fort Worth-based airline, which relied on streaming content to passengers’ phones, laptops, and tablets for many years on domestic flights.
According to Fox News Digital, the airline said that it was a result of changing customer habits.
The airline told Fox News Digital that “customer preferences have changed.” It added that this shift is particularly evident since the COVID-19 epidemic.
Added the carrier, younger travelers in particular expect to have access to larger screens and multiple devices during their trip.
The improvements to satellite internet have also changed what the airlines are able to offer their passengers. American will begin installing Starlink Wi-Fi high-speed on narrowbody aircraft by 2027.
American Airlines reportedly pauses 6 domestic routes due to fuel price pressure tied to the conflict in Iran
Airline said that combining the built-in screens with passengers’ own devices, and faster internet connectivity, could provide a personalized experience in flight, as well as opening “new revenue opportunities and engagement.”
Both new and existing aircraft will be upgraded. The entertainment systems will be installed on newly delivered Airbus A321 or Boeing 737 MAX 10 aircraft, and retrofitted narrowbody aircraft.
American Airlines’ current Airbus A320 aircraft is an exception, and it will not be receiving the new seatback-entertainment systems. The carrier confirmed this to Fox News Digital.
Cabin overhauls go beyond the screens.
American Airlines is increasing its number of premium seats on narrowbody aircraft as it competes with other airlines for passengers willing to pay a higher price for more space or additional amenities.
American narrowbody flights currently have approximately 25% premium seats. This figure is expected to reach approximately 40% within the next few years.
UNITED AIRLINES DROPS MERGER WITH AMERICAN; CEO KIRBY explains WHY
American has already modified its Airbus A319/A320 aircraft in order to include an additional row of seats for first class. The airline plans to add 24 seats to its future Boeing 737 MAX 10 planes. It also intends to modify the A321neo aircraft to include additional first class seating.
Main Cabin Extra is being expanded across the entire narrowbody fleet.
Heather Garboden, American Chief Customer Service Officer said: “We are making the largest investment in onboard experiences in our history.”
As airlines increasingly look to premium passengers and cabin amenities for differentiation, they have made the decision that this is what will happen.
According to American Airlines, the new screen is expected to be among the largest screens available on narrowbody aircrafts in North America. It will offer customized movie and TV recommendations, interactive maps of flights and destination information.
CLICK HERE TO DOWNLOAD FOX NEWS APP
It will take about two years before the first aircraft is delivered with this technology, so passengers can expect to see changes over time rather than an immediate overhaul of the entire fleet.
American Airlines anticipates that the retrofitting program will be complete in the first decade of the 2030s.