AI

Subquadratic Launches A 12-Million-Token AI Model And Says The Wall Is Gone

Published

3 weeks ago

May 6, 2026

A Miami startup called Subquadratic launched its first model on Tuesday with a 12-million-token context window, twelve times the ceiling that every frontier lab has settled near for two years. The company says its proprietary attention architecture scales linearly in compute and memory, runs 52 times faster than dense attention at 1 million tokens, and beats OpenAI’s GPT-5.5 on multi-document recall by nine points. The product launched as an API, a coding agent called SubQ Code, and a deep-research tool called SubQ Search, all running on neoclouds rather than the hyperscalers. Subquadratic has raised $29 million at a $500 million valuation.

The pitch is the kind that surfaces every twelve months in this corner of AI. Magic.dev claimed a 100-million-token window in August 2024, raised more than $500 million, and has produced no public evidence the model is being used outside its own building. Subquadratic’s benchmarks, named experts, and architectural specifics are different. Whether the math survives outside controlled testing is the question every CTO will ask this week, with the company’s launch post detailing SubQ’s SSA architecture and benchmark methodology serving as the opening exhibit.

12 million tokens of context, roughly 9 million words or 120 books in a single prompt.
52x faster than dense attention at 1 million tokens, by the company’s own measurement.
92.1% accuracy on needle-in-a-haystack retrieval at the full 12-million-token length.
$29 million raised at a $500 million valuation from Javier Villamizar, Justin Mateen, and early backers of Anthropic, OpenAI, Stripe, and Brex.

Why The Million-Token Wall Has Held This Long

Quadratic attention has been the bottleneck since the 2017 transformer paper. Doubling the input quadruples the work. The original architecture compares every token to every other token, so a million-token prompt costs a trillion pairwise comparisons. Every workaround the industry has shipped, from retrieval-augmented generation to agentic decomposition to KV-cache offloading, exists to dodge that cost.

The major frontier models have stalled at the same number. Anthropic’s Claude Opus 4.7, Google’s Gemini 3.1 Pro, and OpenAI’s GPT-5.5 all advertise context windows of around one million tokens. None of them work especially well at the upper end. On MRCR v2, the multi-reference recall benchmark labs report, GPT-5.5 leads at 74.0 percent. Opus 4.7 trails at 32.2 percent. Gemini 3.1 Pro lands at 26.3 percent. The advertised window and the usable window are not the same thing.

That gap is what the labs are quietly acknowledging when they pair their long-context offerings with retrieval pipelines. RAG was never a feature; it was a workaround. Subquadratic’s argument is that if attention itself stops scaling quadratically, the workarounds stop being necessary.

Fractured chrome cube bursting open with glowing text fragments representing a 12 million token AI context window.

What Selective Attention Actually Picks

The architecture Subquadratic calls Subquadratic Selective Attention, or SSA, is a learned sparsity mechanism. For any given query token, the model picks which positions in the input matter, conditioned on what the query and keys actually contain. The selection itself does not run quadratic, which is the trap most prior sparse-attention work fell into.

Alex Whedon, Subquadratic’s CTO and a former Meta engineer who ran enterprise AI at TribeAI, described the mechanism in plain terms.

“Sparse attention basically means instead of doing what transformers do, which is if you have 1,000 words, you look at every possible relationship between all 1,000 words, which is 1,000 squared combinations. You realize that only a portion of those actually matter and you only process the portion that matter,” Whedon said in an interview.

The catalogue of prior approaches reads like a graveyard of clever tradeoffs:

Fixed-pattern sparse attention, used in models like Longformer, scales linearly by attending only to a sliding window. It works when relevant information sits nearby. It fails when it doesn’t.
State-space models like Mamba, Mamba-2, RWKV, and RetNet trade dense attention for a recurrent state. The state is lossy. Nvidia’s 8-billion-parameter study found pure Mamba-2 lagged transformers on MMLU and basic phonebook lookup until attention layers were stitched back in.
Hybrid architectures, including Jamba, Kimi Linear, Qwen3-Next, and Nvidia’s Nemotron v3, mix cheap layers with a few dense layers. The dense layers still do quadratic work, so a hybrid that’s three times cheaper at 32K tokens is still three times cheaper at 10 million tokens. The asymptotics never improve.

DeepSeek’s Native Sparse Attention won the ACL 2025 best paper award, and its successor DeepSeek Sparse Attention now ships in DeepSeek’s V3.2-Exp release notes describing the DSA lightning indexer. Sebastian Raschka’s technical breakdown of DeepSeek V3.2 sparse attention notes that DSA reduces complexity from quadratic to linear in the number of selected tokens. The catch, as independent analysts have flagged, is that the lightning indexer that picks those tokens still has to score every query against every key. The selection step itself stays quadratic.

Whedon argues SSA does what DSA tried to do without the indexer trap. “For prompt A, words one and six are going to be important to each other,” he said. “For prompt B, maybe it’s words two and three. It’s different for every single input.” Hybrids, he said, deliver “a scalar benefit.” A pure subquadratic mechanism delivers a scaling-law advantage.

The Benchmarks, And Where They Land

Subquadratic’s own technical paper publishes the headline numbers, with one third-party verification on MRCR v2. On RULER at 128,000 tokens, SubQ scores 97.1 against Opus 4.6’s 94.8. On MRCR v2 the company reports a research result of 83 and a verified production score of 65.9, both ahead of GPT-5.5’s 74. On SWE-Bench Verified, the long-running coding benchmark, SubQ logs 82.4 percent against Opus 4.6’s 81.42 and Gemini 3.1 Pro’s 80.6. At 12 million tokens, where no other frontier model operates at all, SubQ holds 92.1 percent on a needle-in-a-haystack retrieval task. The cost figures are larger than the accuracy figures. Subquadratic’s RULER comparison reports SubQ hit 95 percent at $8 of inference, against Claude Opus’s 94 percent at roughly $2,600.

Benchmark	SubQ	Opus 4.6 / 4.7	GPT-5.5	Gemini 3.1 Pro
RULER 128K	97.1	94.8	not reported	not reported
MRCR v2	83.0 (research)	32.2	74.0	26.3
SWE-Bench Verified	82.4%	81.42%	not reported	80.6%
Needle at 12M	92.1%	not operable	not operable	not operable

Where Subquadratic’s Own Paper Slows Down

The technical paper is unusually candid about its caveats, which is the part most write-ups have skipped. Each model run was performed once. Inference costs at this scale make repeats prohibitive. Standard practice in academic ML is to run benchmarks several times and report median or mean. A single run leaves wider error bars than any of the published deltas.

The SWE-Bench result is also, by the paper’s own description, “harness as much as model.” SWE-Bench scores depend heavily on the agentic scaffolding wrapped around the model: how the harness reads the repository, how it iterates on patches, how it validates tests. A one-point margin over Opus 4.6 may reflect harness design rather than raw model capability.

Whedon also acknowledged that the SubQ model itself is, in his words, “way smaller than the big labs.” Parameter counts have not been disclosed. A subquadratic architecture that performs at frontier scores with fewer parameters is the strongest possible result. A subquadratic architecture that wins because it was carefully tuned for the published benchmarks, while a larger dense competitor runs out of the box, is a much weaker one.

There is also a theoretical ceiling worth flagging. The 2024 paper on fundamental limitations on subquadratic alternatives to transformers proved that for certain reasoning tasks, no truly subquadratic architecture can match dense attention without sacrificing capability. Whether SSA threads that needle, or whether it pays the price on tasks not in the current benchmark suite, will only show up in external use.

What Ships This Week

The launch package is three products. The SubQ API exposes the full 12-million-token window to developers in beta. SubQ Code is a CLI agent that loads an entire repository into a single context call, sidestepping the chunk-and-rerank pipelines most coding agents rely on. SubQ Search is a deep-research tool that runs free at launch as a customer acquisition lever. All three sit on neoclouds, the GPU-specialty providers like CoreWeave and Lambda, rather than AWS or Google Cloud. CEO Justin Dangel told reporters the major hyperscalers are “very expensive.”

The company is not open-sourcing weights. Enterprises that want their own fine-tuned version will get post-training tooling, but the base architecture stays closed. The 50-million-token window target is set for the fourth quarter of 2026. Whether that lands or slips is the first real test of whether SSA scales the way the paper claims. If it does, every retrieval pipeline built in the last three years has a competing answer to consider.

The Magic.dev Shadow Hangs Over Every Big Context Pitch

Subquadratic is not the first startup to claim the ceiling has been broken. The category has a recent and unflattering history that any investor in this round had to weigh.

August 2024, And A Number Nobody Could Test

Magic.dev announced LTM-2-mini’s 100-million-token context window in August 2024, with a claimed 1,000-fold efficiency advantage over Llama 3.1 405B’s attention. The company posted internal benchmarks and raised more than $500 million on the strength of the announcement.

Twenty months later, there is no public evidence of LTM-2-mini being used at scale outside Magic. No third-party benchmark replications. No production customers willing to be named. No follow-up model has shipped publicly. The company’s product page still markets the figure.

What Subquadratic Has To Avoid Repeating

Independent ML analysts have spent the intervening period auditing the broader subquadratic claim. Vladimir Ivanov, in a February 2026 audit of subquadratic attention claims published on LessWrong, surveyed the field and concluded that most reported breakthroughs are best understood as “incremental improvement number 93595 to the transformer architecture” rather than fundamental shifts. Ivanov’s read on Kimi Linear, DeepSeek Sparse Attention, and the state-space family was that each delivered constant-factor speedups while remaining quadratic in their selection or hybrid layers.

Subquadratic’s defense is that SSA is genuinely sparse end-to-end, including the selection mechanism, and that the published benchmarks reflect that. The paper publishes the math. What it cannot publish is the proof the architecture survives at the much larger parameter counts the frontier labs train. That proof comes from external use, or it doesn’t come at all.

$29 Million, A Pivot, And A Speech Model In The Drawer

The $29 million seed at a $500 million valuation is its own data point about the AI funding cycle. The capital came from Javier Villamizar, formerly of SoftBank Vision Fund, and Justin Mateen, the Tinder co-founder who runs JAM Fund. Several individual backers from early Anthropic, OpenAI, Stripe, and Brex rounds also wrote checks. The valuation is roughly 17 times the round size, which is aggressive for a pre-revenue company in any other market and merely typical for AI infrastructure right now.

Subquadratic is a pivot. The company was previously called Aldea and worked on speech models before redirecting to attention architecture. The speech work is shelved. The team that remains is 11 PhD researchers from Meta, Google, Oxford, Cambridge, ByteDance, Adobe, and Microsoft, plus a CEO with five prior companies in health-tech, insurance, and consumer goods, and a CTO who ran enterprise AI at TribeAI after leaving Meta.

Pivots in AI are common right now because the capital is patient and the architecture space is still moving. Whether SSA is the mechanism that finally retires dense attention, or the latest in a long line of constant-factor improvements, is what the next two quarters will tell.

Frequently Asked Questions

Can I Use SubQ Today?

Yes, in beta. The SubQ API opened to developers on May 5, 2026, with the full 12-million-token window exposed. Sign-up runs through subq.ai with an early-access waitlist. SubQ Code is a separate CLI agent built on the same model. SubQ Search runs free at launch. All three sit on neoclouds, not AWS or Google Cloud, so latency and region availability differ from what you’d get from a hyperscaler-hosted model.

How Much Does The SubQ API Cost?

Subquadratic has not posted a public price sheet yet. The company says SubQ runs roughly 50 times cheaper than frontier models at 1 million tokens, with one published RULER comparison showing $8 per benchmark run versus about $2,600 for Claude Opus on the same test. Specific per-token rates will land at general availability. Early access pricing is being negotiated case-by-case for enterprise customers through the company’s sales contact form.

Is SubQ Open Source?

No. Subquadratic has stated it will not release model weights. Enterprises that want their own fine-tuned version will get post-training tooling, but the base architecture stays closed. That puts SubQ in the same camp as Anthropic and OpenAI rather than DeepSeek or Meta. If open weights are a hard requirement, DeepSeek V3.2-Exp on Hugging Face is the closest sparse-attention alternative currently available, though its window tops out around 128K.

When Does The 50-Million-Token Window Launch?

Fourth quarter of 2026, per the company’s published roadmap. The current 12-million-token product is the first commercial release of SSA. The 50-million-token version is described as an extension of the same architecture rather than a new model family. Whether the schedule holds is the most direct test of whether linear scaling holds in production. Subquadratic has not committed to a specific month, and Q4 dates in AI tend to slip.

How Does SubQ Compare To Claude Opus 4.6 On Coding?

Subquadratic’s published numbers show SubQ at 82.4 percent on SWE-Bench Verified, against Opus 4.6 at 81.42. The margin is one point and the paper acknowledges the score is “harness as much as model,” meaning the agentic scaffolding around the model affects the result. For real coding work, the bigger differentiator is SubQ Code’s ability to load a full repository into one context call, which Opus 4.6’s 1-million-token window cannot match.

Will SubQ Work With My Existing IDE?

Through SubQ Code, yes for command-line workflows. Native plugins for Cursor, Cline, or VS Code’s Copilot extension are not yet shipping. You can route the SubQ API into any tool that accepts an OpenAI-compatible endpoint, but the full 12-million-token codebase-loading workflow currently lives only inside the SubQ Code CLI. Plugin integrations are on the public roadmap but not dated, so plan around the CLI for now.

The independent benchmarks will arrive when researchers outside Miami get the API in their own hands and run tests on prompts the company never saw. Until then, the math is on paper and the cautionary tales are documented. Subquadratic has the harder job ahead: producing the receipts.

AI

AAOI Stock Hits All-Time High as AI Optical Spending Surges

Published

3 hours ago

May 26, 2026

Logan Pierce

Applied Optoelectronics, Inc. (NASDAQ: AAOI), a Sugar Land, Texas fiber-optic manufacturer, reported $151.1 million in Q1 2026 revenue on May 7, a 51% jump from the same period a year earlier, with data center sales rising 154% as AI cluster buildouts pushed optical interconnect demand into a new speed tier. Management lifted full-year 2026 guidance to $1.1 billion at a May 13 investor conference, and the stock touched an all-time high of $233.67 the same day after Rosenblatt Securities raised its price target to $220, citing strong Amazon-linked 800G order momentum and upcoming Oracle customer qualifications.

Searches for “AAOI USDT” reflect a broader pattern: crypto-native traders want single-name AI infrastructure exposure without leaving their exchange accounts. The company at that intersection, building the optical transceivers that move data between the GPU racks everyone else discusses, became one of the more volatile story stocks on the Nasdaq in 2026.

Applied Optoelectronics: The Fiber Company Powering AI

AAOI designs and manufactures three families of fiber-optic networking products. Optical transceivers for hyperscale data centers come in 100G, 400G, and 800G configurations, with 1.6-terabit (Tb) modules entering volume production in late 2026. A second line covers hybrid fiber-coaxial (HFC, broadband equipment used by cable operators) products for cable television (CATV) networks. A third covers fiber-to-the-home (FTTH) telecom optics. In Q1 2026, data center revenue crossed CATV for the first time in a meaningful way, the product mix shift that Dr. Thompson Lin, AAOI’s Founder, President and Chief Executive Officer, called central to the company’s growth trajectory.

The company’s role in the AI conversation is structural, not incidental. Every large GPU training cluster needs optical interconnects to move data between chips at the speeds required for parallel processing. As cluster architectures have scaled from hundreds of GPUs to tens of thousands, copper cables hit their physical limits on bandwidth and distance. Optical fiber, which transmits data as pulses of light, handles what copper cannot. AAOI is one of a handful of pure-play vendors serving that supply chain, rather than treating optical as a side line within a broader component portfolio.

What separates AAOI from module assemblers is vertical integration. The company designs and fabricates its own laser diodes, a manufacturing capability that only a handful of competitors match. That integration shortens customer qualification timelines, because hyperscalers can test the complete signal chain at one vendor rather than coordinating across multiple component suppliers. Applied Optoelectronics has expanded its manufacturing footprint in the Houston area to roughly 900,000 square feet across multiple facilities, positioning a significant share of production inside the United States at a time when supply chain geography is a real factor in hyperscaler procurement decisions.

AAOI Applied Optoelectronics AI optical transceiver stock guide for crypto traders.

Why the Optical Layer Is the Structural Bottleneck

TrendForce’s April 2026 AI optical transceiver market analysis placed the global market at $16.5 billion in 2025 and projected expansion to $26 billion in 2026, a 57% single-year increase. The total addressable market estimate for optical transceiver modules was revised upward by 43% and 46% for 2026 and 2027 respectively, driven by hyperscaler capital expenditure plans coming in consistently above consensus. LightCounting, a separate optical research firm, projected the AI cluster-specific transceiver segment alone would double in two years, from $5 billion in 2024 to more than $10 billion in 2026.

The bandwidth demand behind those numbers is architectural. AI clusters are scaling in two directions simultaneously: horizontally, by adding compute nodes, and vertically, by tightening interconnects within the rack. Both trajectories consume optical bandwidth. As GPU cluster sizes increase and rack density rises, passive copper cable solutions hit their physical limits on speed and distance, making optical connectivity a required specification rather than a premium option for new AI infrastructure builds.

Supply constraints are running parallel to the demand acceleration. The bottleneck is not the transceiver module itself but the electro-absorption modulated laser (EML, a chip that converts electrical signals into precise optical pulses) chips and continuous-wave laser diodes inside each unit. Fabricating those components requires precision manufacturing that cannot be scaled in a single product cycle. TrendForce named Applied Optoelectronics alongside Coherent Corp. and Lumentum Holdings, Inc. as vendors that have initiated capacity expansions and technology deployments in direct response to component shortages. AAOI management told attendees at the Needham Technology, Media and Consumer Conference that demand is expected to exceed supply through at least mid-2027.

The technology cycle is also in transition. 800G modules are becoming the standard for new AI cluster deployments this year, while early 1.6T products are entering mass production. TrendForce identifies 2026 through 2027 as the crucial window for vendors to secure design-in at tier-one hyperscale accounts, because success in that qualification cycle determines the revenue trajectory for the following two years. AAOI completed its first volume shipment of 800G modules to a major hyperscale customer in Q1 2026, a qualification milestone rather than a volume event, but qualifications precede volume orders in the optical supply chain.

800G+ share of optical transceiver shipments: 19.5% in 2024, projected above 60% in 2026, per TrendForce
$26 billion: projected global AI optical transceiver market in 2026, up from $16.5 billion in 2025 (TrendForce)
57%: single-year market growth rate as of TrendForce’s April 2026 report, one of the fastest on record for the sector
Mid-2027: AAOI management’s estimate for when demand will continue exceeding available supply capacity

AAOI’s Books: Revenue, Guidance, and the Amazon Signal

First-Quarter Results and What the Numbers Say

Q1 revenue of $151.1 million came in slightly below the Street consensus of roughly $157 million, but the mix told a clearer story. Data center revenue reached $81.4 million, growth of 154% from Q1 2025, and crossed CATV revenue of $66.8 million for the first time in the cycle. The Q1 2026 earnings filing on SEC EDGAR showed GAAP net loss widening to $14.3 million as research and development spending increased and the Texas facility ramped toward capacity. Non-GAAP gross margin came in at 29.1%, easing from 30.6% a year earlier as higher-cost 800G units entered production.

Cash and equivalents stood at $449.4 million at quarter end, following a public equity offering that added approximately $382.4 million. The company also carries $125 million in 2.75% convertible senior notes due 2030 and had $61.7 million in unused borrowing capacity. The balance sheet is funded for the capital expenditure cycle ahead, including the planned Texas expansion and manufacturing equipment for 1.6T product lines.

Metric	Q1 2026 Actual	Q2 2026 Guidance	FY2026 Target
Revenue	$151.1M	$180M to $198M	~$1.1B
Data center revenue	$81.4M	Not disclosed	Not disclosed
CATV revenue	$66.8M	Not disclosed	Not disclosed
Non-GAAP gross margin	29.1%	29% to 30%	Not disclosed
GAAP net income (loss)	($14.3M)	Not disclosed	Not disclosed
Cash on hand	$449.4M	Not disclosed	Not disclosed

The $1.1 Billion Target and Amazon’s Stake

Management raised full-year 2026 revenue guidance to $1.1 billion at the Needham Technology, Media and Consumer Conference on May 13, implying a sharp back-half acceleration. The sequencing is clear: 800G volume ramp with a second hyperscale customer is imminent, 1.6T deliveries are scheduled to begin in late Q3, and the Texas facility adds production output continuously through the year. Q2 guidance of $180 million to $198 million already signals meaningful sequential growth from Q1.

The relationship with Amazon.com, Inc. adds an unusual layer of forward visibility. In March 2025, AAOI issued a customer warrant to a subsidiary of Amazon.com, Inc., granting Amazon the right to purchase up to approximately 7.95 million shares, with vesting tied to Amazon’s purchasing volume over time, potentially covering up to $4 billion in total purchases. The structure means Amazon has a financial incentive aligned with AAOI’s share price, linking the two companies beyond a standard purchase order relationship. Raymond James cited management’s plan to ramp optical transceiver revenue to $1.4 billion by mid-2027 when it raised its target to $160 from $72.50, maintaining an Outperform rating on May 13.

The company locked in more than $324 million in confirmed 800G and 1.6T orders and received a $20.9 million grant from the Texas Semiconductor Innovation Fund in April 2026 to support the Sugar Land facility expansion. That combination of a strategic customer warrant, state manufacturing support, and confirmed order backlog differentiates AAOI from smaller optical vendors competing for the same hyperscale qualifications.

Where the AAOI Thesis Can Break

The bear case is not about the technology or the market size. Both are clearly growing. The question is whether a company burning cash to build manufacturing capacity can execute the production ramp fast enough, at the margins the Street expects, without a competitor closing the qualification gap at one of its major accounts.

Customer concentration: AAOI’s data center revenue flows from a small number of hyperscale accounts. A single large customer adjusting its order cadence can produce double-digit stock moves on earnings day, independent of broader AI infrastructure trends.
Negative free cash flow: Capital expenditure for the Texas expansion kept free cash flow deeply negative through the scaling period. The equity raise addressed near-term liquidity, but sustained profitability is still a future milestone, not a current fact.
Beta of 2.24: The stock amplifies broad market moves by more than double, meaning macro-driven risk-off sessions can produce sharp drawdowns entirely unrelated to AAOI’s own operating results or its customers’ spending plans.
Analyst target dispersion: Price targets ranged from $57.50 at Northland to $220 at Rosenblatt as of mid-May 2026. A spread that wide reflects genuine disagreement about what the ramp is worth at price-to-sales multiples above 20.
Ramp timing risk: B. Riley Securities flagged potential 800G production timing delays into the second half of 2026 even while raising its target to $129 and maintaining a Neutral rating. Q3 and Q4 execution is the central variable that either validates or deflates the annual guidance.

Insider selling in mid-May 2026 drew attention when AAOI executives unloaded a significant block of shares as the stock was near its all-time high. Executive liquidity events at peak prices are common practice, but the timing against still-negative GAAP margins and a price-to-sales ratio above 20 adds to the list of inputs active traders are watching before the July 30, 2026 earnings date.

What “AAOI USDT” Means and How to Trade the Theme

What the USDT Suffix Means for Crypto Traders

Crypto-native traders searching “AAOI USDT” are looking for a Tether-margined perpetual futures contract on AAOI’s price, using the same naming convention as BTCUSDT or ETHUSDT, but referencing a Nasdaq-listed equity rather than a cryptocurrency. The appeal is structural: a single USDT collateral pool, the ability to go long or short with leverage, no fiat onboarding, and consolidated profit and loss alongside existing crypto positions. For a trader already running a multi-asset crypto book, those mechanics represent a real friction reduction compared with opening a separate brokerage account.

AAOI is not currently listed as a perpetual contract on major crypto derivatives platforms. Single-name equity perpetuals require either a licensed stock price oracle, compliance with the securities regulations governing swap instruments, or operation in a jurisdiction where such products are currently permitted. The regulatory landscape around tokenized equities is actively evolving, but as the SEC’s pause on broader tokenized stock access for crypto platforms in May 2026 illustrated, the distance between a single-name equity perp and a compliant crypto exchange product is wider than the search query implies.

Index Futures as the Available Proxy

For traders who want directional exposure to the AI infrastructure theme in a USDT-margined account, Nasdaq 100 (NDX) index futures offer the most accessible available route. The NDX carries heavy weighting toward the hyperscale operators and semiconductor suppliers whose capital expenditure decisions drive demand for AAOI’s products: Amazon.com, Inc., Microsoft Corp., NVIDIA Corp., Alphabet Inc., and Meta Platforms, Inc. together represent a substantial share of the index. Platforms such as Phemex offer USDT-margined NDX futures, making it possible to hold a single USDT account that spans AI-themed index exposure alongside crypto positions.

Feature	AAOI Common Stock	AAOI USDT Perp (Hypothetical)	NDX USDT Futures
Account type	Brokerage account	Crypto exchange	Crypto exchange
Collateral	USD	USDT	USDT
Long / short	Long (cash); short via margin	Both, with leverage	Both, with leverage
Availability (May 2026)	Yes	Not currently listed	Yes
Concentration risk	Single name	Single name	100-name diversified index
Shareholder rights	Yes (voting, proxy)	None	None
Settlement	USD	USDT	USDT

The core trade-off is concentration versus convenience. A direct AAOI position captures the full upside of the 800G ramp and $1.1 billion guidance if the company executes. It also absorbs the full customer concentration risk, the 2.24 beta amplification during risk-off sessions, and any execution shortfalls in Q3 and Q4. An NDX position dampens AAOI-specific upside while surviving an AAOI-specific earnings miss without the same single-session drawdown risk. Which approach fits a given portfolio depends on how much of the thesis is specific to AAOI’s own execution versus how much is simply a bet on AI infrastructure spending continuing at pace.

Frequently Asked Questions

What does Applied Optoelectronics make and why is it considered an AI stock?

Applied Optoelectronics manufactures optical transceivers, laser diodes, and fiber-optic networking products used in hyperscale data centers, cable TV networks, and telecom infrastructure. It qualifies as an AI stock because its data center transceiver products are the optical interconnects that physically move data between GPU clusters inside AI training and inference facilities. Every large-scale AI workload requires high-bandwidth optical links to function, and AAOI is one of a small number of pure-play vendors serving that supply chain with vertically integrated laser and module manufacturing.

What does “AAOI USDT” mean in crypto trading?

AAOI USDT refers to a USDT-margined perpetual futures contract on AAOI’s share price, using the same naming convention as crypto pairs such as BTCUSDT or ETHUSDT. It means a trader would go long or short on AAOI’s price using Tether (USDT, the largest stablecoin by trading volume) as collateral, without a traditional brokerage account. As of May 2026, AAOI is not listed as a perpetual contract on major crypto derivatives platforms.

Is Applied Optoelectronics currently profitable?

Not yet on a GAAP basis. AAOI reported a GAAP net loss of $14.3 million in Q1 2026 and a non-GAAP net loss of $4.9 million, while investing heavily in manufacturing capacity expansion in Texas and in research and development for 800G and 1.6T product lines. On a non-GAAP adjusted EBITDA basis the company reported near breakeven at $966,000 positive in Q1 2026, and Q2 2026 guidance includes the possibility of non-GAAP net income, though GAAP profitability is a later milestone tied to the revenue ramp and margin expansion trajectory.

What is AAOI’s full-year 2026 revenue guidance?

Applied Optoelectronics raised its full-year 2026 revenue guidance to approximately $1.1 billion at the Needham Technology, Media and Consumer Conference on May 13, 2026. Q2 2026 guidance is $180 million to $198 million in revenue with non-GAAP gross margin of 29% to 30%. Management expects sequential revenue growth throughout the year, with the largest acceleration in the second half as 800G volumes ramp with a second hyperscale customer and 1.6T deliveries begin in late Q3.

How can crypto traders get USDT-margined exposure to AI infrastructure themes?

Because AAOI-specific perpetual contracts are not currently listed on crypto exchanges, the most direct available route is USDT-margined Nasdaq 100 (NDX) index futures, which carry heavy weighting toward Amazon, NVIDIA, Microsoft, Alphabet, and Meta, all significant buyers or operators of AI data center infrastructure that drives demand for optical transceivers. Some platforms, including Phemex, offer USDT-margined NDX futures that can sit alongside crypto perpetuals in a single margin account, allowing cross-asset AI infrastructure positioning without fiat onboarding.

AAOI’s next quarterly earnings are scheduled for July 30, 2026. By that date, the back-half ramp driving the $1.1 billion annual target should show up as sequential revenue acceleration and improving gross margins. If both arrive together, the operating data will have caught up with the valuation. If margins compress while revenue grows, the market will want a new answer on profitability timing before extending the multiple further.

Disclaimer: This article is for informational purposes only and does not constitute investment advice or a recommendation to buy, sell, or hold any security or financial instrument. Trading equities, derivatives, and leveraged products involves substantial risk of loss and may not be suitable for all investors. Readers should conduct independent research and consult a qualified financial professional before making any investment decisions. All figures cited are sourced from company filings and industry research accurate as of the publication date and are subject to change.

AI

Claude AI Models Push Investors Past the Chip Trade

Published

3 days ago

May 23, 2026

Logan Pierce

Claude AI models are moving the artificial intelligence (AI) investment debate from chatbot rankings toward security, finance workflow agents, cloud capacity and the hardware needed to run long tasks at scale, after Anthropic said its restricted Mythos Preview system helped partners find more than ten thousand serious vulnerabilities through Project Glasswing and Gartner raised its AI spending outlook.

Portfolio managers can stop asking whether AI remains a theme and start asking which parts of the stack collect the next dollar. The new evidence points to a wider trade: security teams, market-data owners, cloud platforms, memory suppliers, consulting firms and workflow software all sit closer to the revenue line than they did when the contest centered on raw model scores.

The Claude Signal Is No Longer a Chatbot Benchmark

Model news used to trade like a scoreboard. A lab posted a higher benchmark, investors rewarded the chip names, and software vendors promised to add a smarter assistant. That loop still exists, but the AI trade is broadening as models move from answering questions into doing multi-step work inside codebases, spreadsheets and security programs.

The cleanest public marker is the Claude Opus 4.7 release. The company said the model improved on advanced software engineering, long-running coding tasks, vision and enterprise document analysis, while keeping API pricing at $5 per million input tokens and $25 per million output tokens. Application programming interface (API, the developer gateway that lets software call a model) pricing matters because usage can become a recurring cost line instead of a one-off experiment.

The stronger signal may be the model that most customers cannot use. Mythos Preview remains limited because of its computer security capability. That creates a strange investment setup: the restricted system shows what future frontier models may do, while the released system tests whether guardrails, cloud access and paid workflows can turn that power into revenue without blowing up risk budgets.

Claude AI investment shift from chips to enterprise agents.

The Investment Map Has Four New Columns

The usual AI screen starts with chips, cloud and the model lab. This cycle adds workflow packaging, security response and distribution control. That matters because a portfolio built only around training clusters can miss the companies that profit when AI work becomes audited, scheduled and tied to real data.

Signal From the Latest Cycle	Business Asset	Likely Public-Market Exposure	Weak Point
Long-running coding and document reasoning	Developer hours, app modernization and office automation	Cloud platforms, developer tools and IT services	Review cost and accuracy controls
Restricted cyber model	Vulnerability discovery backlog	Cybersecurity vendors, cloud security teams and cyber insurers	Patch triage capacity
Finance agent templates	Packaged analyst workflows	Market-data firms, consulting partners and workflow software	Audit trails and bad source data
Multi-cloud compute contracts	Capacity as distribution	Accelerators, memory, networking, power and data centers	Utilization risk if demand disappoints

That table is why a new model release can move more than one basket. If the model performs, cloud usage rises. If the model performs in finance, data providers get pulled into the workflow. If it performs in security, the industry inherits a patching problem. Each outcome sends dollars to a different vendor class.

Security Turns From Cost Center to Capacity Problem

In cybersecurity, stronger models create two tradable outcomes at once. They make protection spending more urgent, and they threaten parts of the manual testing market by finding flaws faster than human teams can verify them. Anthropic says the bottleneck in Project Glasswing has shifted from finding bugs to verifying, disclosing and patching them.

approximately 50 partners are working with the restricted system through Project Glasswing.
6,202 high or critical estimates were found in open-source projects during the company’s scan, out of 23,019 total vulnerability estimates.
90.6 percent true positive rate was reported among 1,752 assessed findings, with 62.4 percent confirmed as high or critical severity.

The number that should worry security chiefs is smaller: 75 of 530 high or critical disclosed bugs had been patched when the company published the update. That gap is the market. Tools that verify reports, prioritize exposure, route patches and prove remediation become more valuable when discovery volume jumps.

There is a catch. A flood of AI-generated bug reports can bury maintainers and security teams. Investors should be careful with any vendor promising instant protection from model-driven vulnerability discovery. The durable spending may sit with firms that reduce alert noise and document the chain from finding to fix.

Finance Agents Put the Model in Analyst Workflow

The finance release is the clearest product signal for investors because it turns model capability into named jobs. Anthropic introduced finance agent templates for banking and investing work, saying each package combines instructions, governed data connectors and subagents that handle parts of the task such as comparables selection or methodology checks.

The company listed ten ready-to-run finance agents, and said the updates pair best with Opus 4.7, which led Vals AI’s Finance Agent benchmark at 64.37 percent. The agents can run as plugins in Cowork or Code, or as cookbooks for Managed Agents. Claude also works across Excel, PowerPoint and Word, with Outlook listed as coming soon.

Pitch builder creates target lists, runs comparables and drafts meeting books.
Earnings reviewer reads transcripts and filings, then flags thesis changes.
Model builder creates and maintains financial models from filings and data feeds.
Valuation reviewer checks methods against comparables and firm standards.
Month-end closer runs close checklists, prepares journal entries and produces reports.

For public markets, the important point is distribution. A general chatbot lives in an innovation budget. A finance agent embedded in office software, market data and approval flows can land in operating budgets. That gives data owners and consulting partners a stronger claim on AI spending than a thin wrapper around a model.

Compute Commitments Give the Trade Its Toll Roads

The second-order trade needs power, chips and reserved cloud capacity. The model lab captures attention, but the toll-road assets collect rent whenever users call the model, run agents overnight or send documents through compliance review. That is why the infrastructure numbers attached to this release cycle matter.

Cloud Capacity

The Google and Broadcom compute agreement calls for multiple gigawatts of next-generation Tensor Processing Unit capacity starting in 2027. Tensor Processing Units (TPUs, Google’s custom chips for machine learning work) give the company another route around graphics processing units (GPUs, chips used for parallel AI math) and help explain why the AI trade has spread into custom silicon and networking.

Amazon Web Services (AWS, Amazon’s cloud computing arm) has its own claim. Amazon’s latest infrastructure agreement includes a $5 billion investment now, up to an additional $20 billion tied to commercial milestones, a commitment by Anthropic to spend more than $100 billion on AWS technologies over ten years, and up to 5 gigawatts of capacity.

Hardware and Memory

The macro backdrop supports the same read. Gartner’s May AI spending forecast put worldwide AI spending at $2.59 trillion in 2026, up 47 percent from the prior year, with infrastructure accounting for more than 45 percent of the market. That makes capacity a central part of the investment case, not a back-office detail.

The Public-Market Read

IDC’s AI infrastructure tracker note put 2025 AI infrastructure spending at $318 billion and projected $487 billion in 2026. IDC also flagged power generation, grid capacity, memory scarcity and storage constraints as risks. Those constraints are investable, but they are also where margin surprises can appear if capacity arrives late or costs rise faster than usage.

Public Equities Get a Broader Scorecard

The clean winners are no longer limited to one chip supplier or one cloud provider. Hyperscalers gain when model demand turns into reserved capacity, enterprise access and governance controls. Custom silicon suppliers gain if customers want alternatives to GPU scarcity. Memory and networking suppliers gain when inference demand compounds after each new model generation.

Security vendors face the most mixed setup. Better AI can create more demand for scanning, remediation and identity controls, but it can also expose products that depend on expensive manual review. A company that sells triage, proof and patch workflow may benefit more than one that merely adds AI wording to an old scanner.

Software incumbents have a tougher test. If agents live inside spreadsheets, email, documents and code editors, the owner of the workflow can keep the customer relationship. If the model layer pulls work into a separate interface, some traditional software seats lose daily importance. The market will likely reward firms that own data rights, review steps and approvals because those pieces are hard to replace with a prompt box.

The Risk Is Execution, Not Imagination

The bear case is practical. Models can be powerful and still fail in production if they need too much supervision, produce hard-to-audit work or force companies to redesign approvals before savings show up. Gartner noted that enterprises still favor tactical efficiency projects over disruptive change, which means adoption may arrive as a slow budget migration rather than a sudden rewrite of the office.

That is why finance and security are useful tests. Both areas have money, urgency and repetitive work. Both also punish errors. A wrong vulnerability report wastes scarce engineering hours. A wrong financial model can flow into a client deck, audit file or capital decision.

The investment question is therefore less glamorous than the model demo. Can vendors prove verification and control at scale, while the cloud providers deliver enough capacity at the right cost? The answer will decide whether this phase of AI spending lifts many stocks or only the companies closest to usage and governance.

If agent rollouts keep moving into audited workflows, the AI trade broadens from model labs to the companies that carry, secure and govern the work. If verification costs rise faster than productivity, the same releases will look less like a straight line and more like a capex cycle with a margin test.

Disclaimer: This article is for informational purposes only and does not provide investment advice. AI and technology securities carry market, valuation, execution and regulatory risks. Consult a qualified financial professional before making investment decisions. Figures are accurate as of publication.

AI

ChatGPT Tip Helped Drive a Georgia Child Abuse Life Sentence

Published

4 days ago

May 23, 2026

Logan Pierce

The ChatGPT tip in the Corey Hickey case shows how an AI upload can become a child safety report: prosecutors say the tool flagged at least two illegal images, sent a CyberTipline report to the National Center for Missing and Exploited Children, and led Georgia investigators to a phone with more evidence.

On May 21, 2026, a Putnam County jury convicted Hickey, 38, on 31 counts. T. Wright Barksdale III, district attorney for Georgia’s Ocmulgee Judicial Circuit, said the judge imposed two life sentences without parole, three life sentences with parole eligibility and 220 years in prison to run consecutively.

The Tip That Started a Criminal Case

The public record begins with a cyber tip. The Georgia Bureau of Investigation (GBI, the state agency that handled the cybercrime inquiry) said it began investigating Hickey’s online activity in October 2025 after receiving a report from the center about possible production, possession and distribution of child sexual abuse material (CSAM, illegal material depicting child sexual exploitation). The Georgia Bureau of Investigation arrest notice says Hickey was charged with three counts of sexual exploitation of children and booked into the Putnam County Jail.

Prosecutors later said the case widened after agents confirmed the images were taken in Putnam County and seized Hickey’s phone. A forensic review of that device uncovered numerous videos, leading to additional warrants for rape, aggravated assault and child molestation. The original online alert became the doorway to a local evidence case.

The indictment shows the shift. Prosecutors said jurors convicted Hickey on three counts of rape of a child under 10, four counts of aggravated child molestation, three counts of child molestation and 21 counts of sexual exploitation of children. That mix matters because it ties the online upload to alleged in-person abuse, device evidence and a full trial record.

ChatGPT NCMEC child safety report in Georgia abuse case.

The AI Report Chain Has Four Hands

The technology piece can sound instant: upload, flag, report, arrest. The working chain is slower and more human. The case moved through a four-hand relay involving a company safety system, a national clearinghouse, state investigators and local prosecutors.

The CyberTipline report process at NCMEC says staff review tips and work to find a potential location so reports can be made available to the proper law enforcement agency. That routing function is why the same pipeline can serve a social network, a cloud service, a messaging app or an AI product.

Actor	Role in a CyberTip Case	Publicly Confirmed in This Case
OpenAI	Detects and reports prohibited uploads or requests involving CSAM.	Prosecutors said at least two images uploaded to the AI tool were flagged and sent onward.
NCMEC CyberTipline	Reviews reports and routes them to the law enforcement agency best placed to respond.	GBI said its inquiry began after a CyberTipline report from the center.
GBI CEACC Unit	Investigates online child exploitation leads with local partners.	GBI said its Child Exploitation and Computer Crimes Unit opened the inquiry and made the arrest.
Ocmulgee Judicial Circuit DA	Turns investigative findings into charges and trial proof.	Prosecutors said the jury returned guilty verdicts across the full indictment.

That division of labor is easy to miss. The company can detect and report. The center can triage and route. Investigators still need warrants, devices, interviews and forensic work before a prosecutor can ask a jury for a conviction.

The Scale Behind One Tip

One Georgia prosecution landed inside a reporting system that now handles industrial-scale volume. The latest CyberTipline data from NCMEC says the service received 21.3 million reports in 2025, and provider reports included 61.8 million images, videos and other files.

21.3 million reports reached the CyberTipline in 2025.
61.8 million files were attached to electronic service provider reports.
107,817 reports were submitted by OpenAI to the center from July to December 2025, according to OpenAI child safety reporting totals.

The data also shows why generative AI has become a child safety category of its own. NCMEC said 2025 reports included 1.5 million with a generative AI nexus, though it noted that more than 1.1 million came from Amazon AI Services detections tied to potential CSAM inside training datasets and did not include actionable offender or victim information.

That caveat is important. A report can describe a known image, a new upload, a cloud storage hit, an attempted generation or a training-data detection. Law enforcement value depends on what comes with the alert: account details, timestamps, location signals, preserved files and enough context to identify a suspect or protect a child.

Detection Happens Before a Courtroom

The case answers a common tech question in blunt terms: uploads to AI services are subject to safety checks when they involve suspected child exploitation. OpenAI says its rules bar users from exploiting, endangering or sexualizing anyone under 18, and that its systems monitor for child safety violations.

Our Child Safety Team reports all instances of CSAM, including uploads and requests, to NCMEC

That line appears in OpenAI’s child exploitation safety post, which also says the company uses hash matching for known material and Thorn’s CSAM classifier for potentially novel material. Thorn is a child safety technology nonprofit whose tools are used by platforms trying to identify abuse material at scale.

AI-specific abuse adds another layer. The same service can receive an old illegal image, a newly produced file, an attempt to generate abusive content or a request asking the model to describe an uploaded file. Detection has to sort those cases quickly, then hand the highest-risk ones to a human process that can preserve evidence and escalate urgent threats.

Prosecutors Still Needed Local Evidence

A CyberTipline report can start the knock on the door, but prosecutors still have to prove crimes under state law. In Hickey’s case, the proof described by prosecutors moved from the AI upload to physical jurisdiction, a seized phone and a broader digital review.

Location: investigators confirmed the images tied to the tip were taken in Putnam County.
Device evidence: Hickey’s phone was seized after his arrest and reviewed forensically.
Cloud evidence: prosecutors said more than 400 additional images were found on his Google Drive.
Charging decision: the case grew from three initial sexual exploitation charges into a 31-count indictment.

Prosecutors said law enforcement found over 4,800 images and videos of child abuse material on the phone. That volume did not replace the jury’s job. It gave prosecutors a body of local evidence that could be tested in court, tied to a device and placed beside the testimony and forensic work needed for conviction.

The Google Drive detail also shows why these cases rarely stay inside one product. A single tip can point to a handset, then to cloud storage, then to accounts, timestamps and other services. For investigators, the first alert is often a map fragment, not the whole map.

The Privacy Trade-Off Sits in the Same File

The safety win carries a privacy tension that deserves plain language. AI companies invite users to upload files, images and videos for help, but some categories of content trigger monitoring, account action and legal reporting. The boundary is drawn around suspected harm, yet the mechanism still depends on scanning user-provided material.

The company’s U.S. privacy policy for uploaded content says user content can include prompts and uploaded files, images, audio and video. It also says personal data may be used to comply with legal obligations and protect the rights, privacy, safety or property of users, the company or third parties.

That is the bargain every large platform now has to defend. Weak detection leaves children exposed and lets offenders use mainstream services as cover. Overbroad systems can create false alarms, chill lawful use and put sensitive material in front of reviewers. The hard standard is not just catching more. It is catching better, with secure handling, clear escalation and records a court can understand.

The Alert Is Only the First Rescue

The deepest fact in the Georgia case is that an online report surfaced alleged offline abuse. Prosecutors said the victim was a 7-year-old child known to Hickey, and the uploads gave investigators a path into a case that was larger than the two images that triggered the alert.

NCMEC says its staff review tips, seek a potential location and make reports available to appropriate law enforcement. The GBI said its inquiry was part of the Internet Crimes Against Children (ICAC, a U.S. Department of Justice-backed task force program) effort housed with the state agency’s child exploitation unit. In practice, that means child safety work moves across company systems, nonprofit analysts, state investigators and county courtrooms.

For readers who suspect online child exploitation, the practical route remains direct reporting through the CyberTipline or local law enforcement rather than confronting a suspect or circulating material. The Georgia case shows why: the evidence has to be preserved, routed and handled by people trained to protect both the child and the case.

A model can refuse an upload in milliseconds, but the child is protected only when the report reaches someone with a badge.