EMERGING TECH RESEARCH

Artificial Intelligence
& Machine Learning
Report
VC trends and emerging opportunities

Q4
2023
Contents
Vertical update 3 Institutional Research Group
Analysis
Q4 2023 timeline 4
Brendan Burke Senior Analyst, Emerging Technology
AI & ML landscape 5
[email protected]

AI & ML VC ecosystem market map 6 [email protected]

VC activity 7 Data

Emerging opportunities 15 Matthew Nacionales Data Analyst

AI datacenters 16 Publishing
Domain-specific foundation models 20 Report designed by Megan Woodard and Julia Midkiff

Published on February 27, 2024

Local LLMs 24

Select company highlights 27

Anthropic 28
For previous updates as well as our complete AI & ML
Databricks 31 research, please see the designated analyst workspace
on the PitchBook Platform.

Q4 2023 Artificial Intelligence & Machine Learning Report CONFIDENTIAL. NOT FOR REDISTRIBUTION. PG 2
Vertical update
The financial forecasts from our Q3 2023 AI & ML public valuation guide and comp sheet are
Q4 2023 Morningstar/PitchBook unicorn industry vertical indexes
beginning to materialize in blowout earnings guidance for a range of horizontal software giants and
total return*
traction for AI service providers, emboldening the ecosystem of startups supporting tech leaders.
Thus far in Q1 2024, standout earnings releases with AI exposure included Nvidia, Microsoft,
Artificial intelligence
ServiceNow, Arm, Meta, Palantir, SK Hynix, IBM, and Qualcomm. These results demonstrate
enterprise efforts to integrate Nvidia GPUs into their datacenters, centralize data for AI retrieval,
Fintech
and deploy AI-integrated applications in IT and revenue operations departments. Not all public
companies benefit equally, yet leadership in innovation is translating to outstanding results. These E-commerce
financial results are diffusing to a handful of startups offering models to cloud giants including
Anthropic and OpenAI, both of whom have robust forecasts for revenue growth in 2024. Cybersecurity

SaaS
These results build on heightened expectations from Q4 that improved valuation expectations
for all AI unicorns, not just foundation model startups. The Morningstar PitchBook Global Unicorn Healthtech
Industry Vertical Indexes mark unicorn valuations to a proprietary model via a combination of
comparable public and private company valuations and precedent VC deals. Both outstanding Mobility tech
unicorn deals and comparable company appreciation led the AI index to outperform all other
Supply chain
verticals in Q4 with a 14.1% total return, even as cybersecurity led the field for the year. Public
comparable company valuations in semiconductors and pure-play AI core software drove
Climate tech
unicorn valuations higher, as laid out in our Q4 2023 AI & ML Public Comp Sheet and Valuation
guide. Valuation growth for unicorns including Anthropic, AI21 Labs, Neuralink, and Shield AI also Biopharma
contributed to this outperformance. Q1 2024 share price gains should lead to continued leadership
0% 2% 4% 6% 8% 10% 12% 14% 16%
for the AI index.

Source: Morningstar and PitchBook • Geography: Global • *As of December 31, 2023
Note: The Agtech index was excluded for scale.

Q4 2023 Artificial Intelligence & Machine Learning Report CONFIDENTIAL. NOT FOR REDISTRIBUTION. PG 3
 Q4 2023 timeline Q4 VC deal activity

October 3 November 21 December 9

Anthropic raises a $2.0 billion round AI21 Labs raises a $208.0 million late-stage round. The European Parliament reaches 1,665
from investors including Alphabet. The The company’s foundation models stand out in recent a provisional agreement on the total deals
corporate partnership diversifies the testing for their calibration, which refers to awareness Artificial Intelligence Act. The
$22.3B
VC deal

VC deal
company’s cloud partnerships after of their limitations and disclosure of uncertainty. Act limits the use of AI in harmful

News
receiving funding from Amazon in Q3. AI21 Labs has developed AI21 Studio, a pay-as-you-go applications and creates criteria for total VC raised
developer platform for building custom text-based publication of models deemed high-
business apps based on AI21’s LLMs. As a result, AI21
Labs competes with foundation model labs, including
risk in terms of size and potential
influence on elections.
11.0%
Cohere and Anthropic, as well as marketing-focused
deal value growth QoQ
vendors such as Jasper, Regie, and Typeface.

2023 YTD summary

Oct 1

Nov 1

Dec 1

Dec 31
October 24 December 5 7,238
Perplexity raises $73.6 million of Mistral raises a $433.9 million early-stage round at a $2.2 total deals
Series B venture funding in a deal led billion post-money valuation led by Andreessen Horowitz.
by IVP. The company has developed In open-source adoption, Mistral’s large language model
$90.9B

VC deal
VC deal

a search engine that is competitive has become a top-10 text generation model on Hugging
total VC raised
with Google and Microsoft’s Bing Face, surpassing 500,000 downloads. By comparison,
since being founded in 2022. MosaicML garnered a $1.3 billion valuation from Databricks,
and its LLM has only surpassed 100,000 downloads on the -4.7%
same platform. Mistral AI’s models, including Mistral 7B deal value growth YoY
and Mixtral 8x7B, differentiate themselves from OpenAI’s
GPT-4 and Anthropic’s Claude 2 in cost and efficiency.

Q4 2023 Artificial Intelligence & Machine Learning Report CONFIDENTIAL. NOT FOR REDISTRIBUTION. PG 4
AI & ML landscape
Horizontal platforms

Vertical applications
$
Semiconductors

Autonomous machines

Q4 2023 Artificial Intelligence & Machine Learning Report CONFIDENTIAL. NOT FOR REDISTRIBUTION. PG 5
AI & ML VC ecosystem market map
This market map is an overview of venture-backed or growth-stage companies that have received venture capital or other notable private investments. Click to view the full map on the PitchBook Platform.

Horizontal platforms Vertical applications

AI automation platforms Consumer Healthcare IT

Scientia
Technologies

AI core Financial services Industrial Transportation

Foundation models

Semiconductors Autonomous machines

Computer vision Datacenter Intelligent sensors & devices Autonomous vehicles

Natural language technology Edge AI software Processor design Intelligent robotics

Q4 2023 Artificial Intelligence & Machine Learning Report CONFIDENTIAL. NOT FOR REDISTRIBUTION. PG 6
VC activity AI & ML VC deal activity

VC activity has remained consistent throughout the VC downturn over the past six quarters, 9,645
8,968
leaving aside the tech giant megadeals that commit upfront funds for future cloud purchases. We
7,253 7,238
6,283 6,811
tracked $22.3 billion invested in Q4, including an outlier $2.0 billion deal size for Anthropic, led
by Alphabet. Median valuations remain above the market midpoint while equity sold falls lower,
demonstrating founder-friendly conditions. Deal count has remained sharply lower than pre-
$61.9 $60.5 $74.4 $143.0 $96.6 $90.9
downturn levels, even given generative AI (GenAI) excitement over the past five quarters. These
totals would assuredly be much lower without tech giants’ involvement in GenAI deals, with GenAI 2018 2019 2020 2021 2022 2023*
leaders raising $6.0 billion in Q4 across only 194 deals. The momentum in horizontal platforms Deal value ($B) Deal count
led the segment to set a VC record in 2023 with $33.0 billion raised, even as vertical applications
regressed to 2020’s level. Source: PitchBook • Geography: Global • *As of December 31, 2023

Q3’s bump in VC exit M&A did not sustain in Q4, as both deal value and count fell. We only
tracked $2.7 billion in disclosed deal value in Q4, which will likely settle as the lowest quarter since
AI & ML VC exit value ($B) by type
Q1 2019. Tech giants remained dormant in M&A given their focus on partnerships with leading
LLM startups. Exceptions included AMD’s acquisition of Nod.AI in machine learning operations
(MLOps), IBM’s acquisition of Manta in database management, and ServiceNow’s acquisition of $300
Buyout
UltimateSuite in predictive analytics. An upcoming IPO for semiconductor startup Astera Labs
Acquisition
promises to revive deal value in Q1 or Q2 2024. Impressive results for AI companies in public $200
Public listing
markets should encourage further listings, although our review of the IPO pipeline for this year
does not heavily feature AI companies given their ability to stay private. $100

$0
2018 2019 2020 2021 2022 2023*

Source: PitchBook • Geography: Global • *As of December 31, 2023

Q4 2023 Artificial Intelligence & Machine Learning Report CONFIDENTIAL. NOT FOR REDISTRIBUTION. PG 7
VC ACTIVITY

Median AI & ML VC deal value ($M) by stage Median AI & ML VC pre-money valuation ($M) by stage

$35 $350

$30 $300

$240.1
$23.3
$25 $250

$20 $200

$15 $14.7 $150

$110.0
$10 $8.3 $100 $38.6
$7.0 $19.7
$5 $3.4 $50 $8.0
$3.1 $38.0
$25.3
$2.1 $10.0
$1.7
$0 $0
2018 2019 2020 2021 2022 2023* 2018 2019 2020 2021 2022 2023*
Pre-seed/seed Early-stage VC Late-stage VC Venture growth Pre-seed/seed Early-stage VC Late-stage VC Venture growth

Source: PitchBook • Geography: Global • *As of December 31, 2023 Source: PitchBook • Geography: Global • *As of December 31, 2023

Q4 2023 Artificial Intelligence & Machine Learning Report CONFIDENTIAL. NOT FOR REDISTRIBUTION. PG 8
VC ACTIVITY

AI & ML VC deal value ($M) by vertical application AI & ML VC exit count by type

$22,000 700
Buyout
$20,000 Acquisition
600
$18,000 Public listing

$16,000
500
$14,000

$12,000 400

$10,000
300
$8,000

$6,000
200
$4,000

$2,000 100

$0
2019 2020 2021 2022 2023* 0
Consumer Financial services Healthcare IT Industrial AI Transportation 2018 2019 2020 2021 2022 2023*

Source: PitchBook • Geography: Global • *As of December 31, 2023 Source: PitchBook • Geography: Global • *As of December 31, 2023

Q4 2023 Artificial Intelligence & Machine Learning Report CONFIDENTIAL. NOT FOR REDISTRIBUTION. PG 9
VC ACTIVITY

Key Q4 AI & ML early-stage VC deals by deal value ($M)*

Deal value Valuation step-

Company Close date (2023) Subsegment, category Lead investor(s)
($M) up (post to pre)

Aleph Alpha November 6 AI core, foundation models $486.2 Bosch Ventures, Ipai, Schwarz Unternehmenskommunikation N/A

Mistral AI December 5 AI core, foundation models $433.9 Andreessen Horowitz, General Catalyst, Lightspeed Venture Partners N/A

01.AI December 14 AI in IT, productivity optimization $200.0 N/A N/A

Didi Autonomous Driving October 13 Autonomous vehicles, autonomous vehicle software $149.0 GAC Capital N/A

xAI December 5 N/A $134.7 N/A N/A

together.ai November 2 AI core, foundation models $102.5 Emergence (Financial Services), Kleiner Perkins, Nvidia 4.6x

Stability AI October 1 AI core, foundation models $86.0 Intel, Sound Ventures, Millennia Capital N/A

Harvey December 19 AI in IT, legal automation $80.0 Elad Gil, Kleiner Perkins 4.2x

Perplexity November 6 Consumer AI, media & entertainment $73.6 IVP 3.6x

Hayden AI December 21 Autonomous vehicles, autonomous vehicle software $53.0 Drawdown Fund 1.3x

Source: PitchBook • Geography: Global • *As of December 31, 2023

Q4 2023 Artificial Intelligence & Machine Learning Report CONFIDENTIAL. NOT FOR REDISTRIBUTION. PG 10
VC ACTIVITY

Key Q4 AI & ML late-stage VC deals by deal value ($M)*

Deal value Valuation step-

Company Close date (2023) Subsegment, category Lead investor(s)
($M) up (post to pre)

Anthropic October 27 AI core, foundation models $2,000.0 Alphabet N/A

Metropolis October 5 Transportation, ADAs $1,700.0 3L Capital, Eldridge Industries N/A

Ramp December 6 AI in financial services, wealth management $339.0 Sands Capital, Thrive Capital 0.7x

Neuralink November 22 Intelligent sensors & devices $323.2 Founders Fund 1.5x

Verkada October 9 Intelligent sensors & devices $305.0 Alkeon Capital Management, Linse Capital 1.9x

Lambda October 5 Chips, GPU cloud $300.0 Mercato Partners 7.4x

AI21 Labs November 21 AI core, foundation models $208.0 Pitango Venture Capital, SCB 10X, Samsung NEXT Ventures, Walden Catalyst, b2venture 2.1x

Lightmatter December 19 Chips $155.0 Aliya Capital Partners, GV, SIP Global Partners, Viking Global Investors 1.6x

Crusoe December 5 Chips, GPU cloud $143.7 N/A N/A

Pony.ai October 24 Autonomous vehicles, autonomous vehicle software $100.0 NEOM Investment Fund 1.0x

Source: PitchBook • Geography: Global • *As of December 31, 2023

Q4 2023 Artificial Intelligence & Machine Learning Report CONFIDENTIAL. NOT FOR REDISTRIBUTION. PG 11
VC ACTIVITY

Key Q4 AI & ML VC exits by exit value ($M)*

Company Close date (2023) Subsegment, category Exit value ($M) Post-money valuation ($M) Exit type Acquirers(s)

Clearpath Robotics October 2 Intelligent robotics, industrial robots $454.5 $454.5 Acquisition Rockwell Automation

LeddarTech December 22 Intelligent sensors & devices $259.0 N/A Public listing Prospector Capital

Hivestack December 12 Consumer AI, ad targeting $100.0 $100.0 Acquisition Perion Network

Ponder October 24 AI automation platforms, database management $46.0 $46.0 Acquisition Snowflake

Tessian December 19 AI in IT, information security automation N/A N/A Buyout Proofpoint, Thoma Bravo

UltimateSuite December 18 AI automation platforms, predictive analytics N/A N/A Acquisition ServiceNow

Heap December 7 AI automation platforms, predictive analytics N/A N/A Acquisition ContentSquare

WireWheel December 4 AI in IT, information security automation N/A N/A Acquisition Osano

Nod.AI October 19 AI core, model development tools N/A N/A Acquisition Advanced Micro Devices

Sisu Data October 16 AI automation platforms, predictive analytics N/A N/A Acquisition Snowflake

Source: PitchBook • Geography: Global • *As of December 31, 2023

Q4 2023 Artificial Intelligence & Machine Learning Report CONFIDENTIAL. NOT FOR REDISTRIBUTION. PG 12
VC ACTIVITY

Top strategic acquirers of AI & ML companies since 2017* Top VC investors in AI & ML companies since 2017*

Pre-seed/ Early- Late-stage Venture Investor

Investor Deal count Investor type Investor Deal count
seed stage VC VC growth type

Accenture 21 Corporation Alumni Ventures 400 163 125 97 15 VC

Apple 21 Corporation Sequoia Capital 307 82 117 78 30 VC

Microsoft 14 Corporation HongShan 275 30 171 59 15 VC

Cisco Systems 13 Corporation 500 Global 260 123 78 50 9 VC

Meta 13 Corporation SOSV 254 130 61 53 10 VC

ServiceNow 11 Corporation Andreessen

248 56 93 69 30 VC
Horowitz

International Business Machines 10 Corporation Right Side Capital

248 119 111 17 1 VC
Management
DataRobot 9 VC-backed company
Enterprise Ireland 236 90 87 53 6 VC
Intel 9 Corporation
Accel 235 49 86 84 16 VC
Snowflake 8 Corporation
Antler 234 206 28 0 0 VC
Source: PitchBook • Geography: Global • *As of December 31, 2023
Source: PitchBook • Geography: Global • *As of December 31, 2023

Q4 2023 Artificial Intelligence & Machine Learning Report CONFIDENTIAL. NOT FOR REDISTRIBUTION. PG 13
VC ACTIVITY

Top VC- and PE-backed AI & ML companies by total VC raised to date*

Company VC ($M) raised to date Segment Subsegment IPO probability M&A probability No exit probability

OpenAI $10,310.0 Horizontal platforms AI core 34% 64% 2%

Anthropic $7,254.0 Horizontal platforms AI core 93% 4% 3%

Waymo $5,500.0 Autonomous machines Autonomous vehicles 40% 58% 2%

Databricks $4,181.9 Horizontal platforms AI core 91% 7% 2%

Relativity $2,383.5 Vertical applications Industrial AI 78% 6% 16%

Anduril $2,315.1 Autonomous machines Intelligent robotics 68% 30% 2%

Nuro $2,132.0 Autonomous machines Intelligent robotics 74% 24% 2%

JD Digits $2,127.9 Vertical applications Financial services N/A N/A N/A

Horizon Robotics $1,920.0 AI & ML semiconductors Processor design N/A N/A N/A

Inflection $1,565.0 Horizontal platforms AI core 27% 71% 2%

Source: PitchBook • Geography: Global • *As of December 31, 2023

Note: Probability data is based on PitchBook VC Exit Predictor methodology.

Q4 2023 Artificial Intelligence & Machine Learning Report CONFIDENTIAL. NOT FOR REDISTRIBUTION. PG 14
Emerging opportunities
AI datacenters Domain-specific foundation models

Today’s datacenter investments in Newly open-sourced datasets encourage

generative AI will take until 2027 to pay development of new model architectures.
off with SaaS-style margins, assuming
continued high growth.

Local LLMs

Open-source models serving projects by

Ollama, Llama.cpp, and Oobabooga have
escalated rapidly in developer adoption
since Q2 2023.
AI datacenters
Overview
Comparison of generative AI hardware and software spending ($B)
AI datacenters distinguish themselves from traditional datacenters through their specialized
infrastructure and operations designed to meet the rigorous demands of AI workloads. They are
$75
equipped with substantial computing resources, including high-performance servers, storage,
$70
and networking, as well as specialized accelerators, to efficiently process AI tasks. Due to the
$65
intense power and cooling needs required for high-performance GPUs, AI datacenters necessitate
$60
enhanced power connections and innovative cooling methods to manage the significantly
$55
higher power consumption and heat output. These centers feature a higher power density, often
$50
exceeding 60 kilowatts per rack, far surpassing traditional setups. Their location can be more
$45
flexible, allowing for cost savings on real estate, and they embody a greater level of complexity and
$40
planning to ensure reliability and mitigate potential failures.
$35
$30
Spending on hardware for AI datacenters currently exceeds the revenue for software applications
$25
built on top of them and may take three years to pay off capital investments and five years to
$20
generate high SaaS-style margins. According to IDC data, $8.0 billion was spent on generative
$15
AI processors, storage, and networking in 2023, producing $2.1 billion in cloud revenue and $4.5
$10
billion in application sales.1 We assume the vast majority of hardware spending came from cloud
$5
datacenter owners such as Microsoft, Amazon, and Google. While cloud and application spending
$0
will exceed datacenter hardware spending in 2024, the cost of goods sold will keep generative AI
2023 2024 2025 2026 2027
investments cash-flow-negative for several years, with high margins not likely to be generated
Generative AI cloud services & application spending Generative AI data center hardware
until 2027. This low-margin period will challenge some lesser-funded datacenter providers to
retain customers and continue offering GPUs at low cost.
Source: IDC • Geography: Global

1: “Worldwide Core IT Spending for GenAI Forecast, 2023–2027: GenAI Is Triggering Hyper-Expansion of AI Spending” IDC, December 20, 2023.

Q4 2023 Artificial Intelligence & Machine Learning Report CONFIDENTIAL. NOT FOR REDISTRIBUTION. PG 16
AI DATACENTERS

GPU cloud hosts market map

Market direction AI datacenter providers
Hyperscalers
Startup cloud providers are undercutting hyperscalers on cost and standing out in supply of
advanced semiconductors. According to hourly on-demand pricing, startups are offering 50%-
70% cost savings on GPU hours for advanced Nvidia A100s and offering unique access to the
latest H100 chips. 2 Leading startup GPU cloud provider Lambda has built the largest cluster of Diversified colocation hosts with AI capabilities
H100 chips of all public clouds, exceeding Google and Oracle. 3 Nvidia’s H100 chips are specially
designed for transformer model training, encouraging use by leading training labs. Nvidia has since
announced the H200 with enhanced memory that is exclusively offered by some AI specialists
AI-focused GPU cloud hosts
including Lambda. Nvidia is developing its own cloud services via colocation partners instead
of building its own cloud datacenters. Enabling developers to cheaply build on independent
GPU clouds offers the possibility to scale revenue non-linearly if startup products gain scale. At
present, startups remain core customers for AI cloud specialists, creating risk of customer churn.

GPU rental services

While hyperscalers dominate cloud computing, an ecosystem of colocation providers for
enterprises developing local services can produce promising investment opportunities. Specialty
cloud providers have carved out a $4.6 billion market from the nearly $150 billion internet-as-
a-service market, more than 90% of which accrues to US-based hyperscalers and China cloud
giants.4 This market segment is large enough to support large VC-backed companies. Specialty
cloud providers stand out not only for AI chip availability but local presence, multicloud support,
edge computing, and support for multiple types of legacy hardware. Additionally, Equinix has
launched a private cloud service for Nvidia SuperPODs, further strengthening their years-long

2: “Cloud GPU Comparison,” Jolt, n.d., accessed February 8, 2024.

3: “State of AI Report Compute Index,” Air Street Capital and Zeta Alpha, n.d., accessed February 13, 2024.
4: “Market Share: Enterprise Public Cloud Services, Worldwide, 2022,” Gartner, July 25, 2023.

Q4 2023 Artificial Intelligence & Machine Learning Report CONFIDENTIAL. NOT FOR REDISTRIBUTION. PG 17
AI DATACENTERS

collaboration. This partnership is part of Nvidia’s broader efforts to integrate its AI technology At the early stage, new cloud providers are coming online. In Q4, RunPod raised a $18.5 million
into specific domains, including robotics and datacenter architecture. Startup vendors including seed round to bring a scalable GPU cloud to market with serverless software. The company was
CoreWeave, Lambda, Together.ai, and RunPod figure prominently in this landscape. founded in 2022 and joined the AI Grant program run by leading AI investors Nat Friedman and
Daniel Gross. The company is primarily working with AI startups, and we believe the company’s

Trending startups business model involves rental of GPUs from cryptominers. The company offers the second-
lowest hourly prices on the market for Nvidia 40 GB A100 and A600 instances while making H100s
In Q4, GPU cloud startup Lambda achieved an outlier valuation step-up of 7.4x to a $1.3 billion available. The company benefits from the current disillusionment with Docker containers with
pre-money valuation. The valuation is justified by the company’s revenue, which the company serverless deployment while being closely involved with the Silicon Valley startup community.
forecast to reach $250.0 million in 2023. 5 The company expects further 100% growth in 2024. The GPU rental has proven to be a difficult business model for startups, as evidenced by early GenAI
company stands out not only for GPU availability via its GPU Cloud, founded in 2018, but also AI innovator Banana’s decision to leave the market. Other alternatives to RunPod include Replicate
compiler software that leverages popular open-source model training frameworks with a simple and Modal Labs.
wrapper. Lambda’s client roster includes GenAI unicorns such as Imbue, Picsart, and Writer. Imbue
chose to pursue independent partnerships with companies including Dell and Lambda instead of
partnering with a hyperscaler. Because of the limited supply of GPUs, Lambda has become a target
for growth-stage companies yet may face churn as the competitive landscape shifts in favor of
large companies.

5: “Lambda Labs Nears $300 Million Financing,” The Information, Kate Clark, October 10, 2023.

Q4 2023 Artificial Intelligence & Machine Learning Report CONFIDENTIAL. NOT FOR REDISTRIBUTION. PG 18
AI DATACENTERS

Key Q4 recent AI datacenter VC deals ($M)*

Company Close date (2023) Deal value ($M) Post-money valuation ($M) Lead investor(s)

RunPod December 4 $18.5 $68.5 N/A

Lambda October 5 $300.0 $1,550.0 Mercato Partners

Kao Data October 5 $60.3 $423.7 Legal & General Group, Infratil

Genesis Cloud October 1 $20.5 N/A N/A

CoreWeave May 31 $421.0 $2,521.0 Magnetar Capital, Nvidia

Source: PitchBook • Geography: Global • *As of December 31, 2023

Q4 2023 Artificial Intelligence & Machine Learning Report CONFIDENTIAL. NOT FOR REDISTRIBUTION. PG 19
Domain-specific
foundation models Generative AI emerging space VC deal count by segment

Overview
225

Domain-specific foundation models (DFMs) leverage cutting-edge model architectures with

200
unique datasets to identify novel intelligence not found in conventional large language models.
Foundation models differentiate based on the usage of self-attention mechanisms to emphasize 175
specific characteristics of a dataset. DFMs can differ from LLM architectures based on their
datasets and balance between encoder and decoder layers. Particularly, multimodal datasets with 150

both images and text can require customized attention mechanisms and fine-tuning strategies.
125
Some pre-existing AI models may be relabeled as foundation models, but we believe investors can
differentiate based on the underlying architecture and datasets involved. 100

Vertical applications have been secondary beneficiaries of the GenAI wave and are likely to 75
produce some of the largest businesses through DFMs. In 2023, vertical application deal count
50
grew more than any other segment in generative AI, with 112.9% growth to 132 deals—second
only to AI core software. Vertical applications surpassed natural language interfaces, a segment 25
that includes major use cases in customer support, sales & marketing, and search. Much of
this growth owed to the emergence of new use cases that previously did not take advantage 0
of GenAI, including corporate strategy, fintech, industrial, and legal. Early-stage startups lead 2019 2020 2021 2022 2023*

the development of standalone businesses built on AI in these domains. If the field of GenAI Audio Healthcare Code Visual media Natural language interfaces Vertical applications AI core

progresses like supervised machine learning did, then large AI-native businesses will be created in
each large end market. Source: PitchBook • Geography: Global • *As of August 18, 2023
Note: A list of companies in each segment can be found in our
Generative AI emerging space market map Q1 2024.

Q4 2023 Artificial Intelligence & Machine Learning Report CONFIDENTIAL. NOT FOR REDISTRIBUTION. PG 20
DOMAIN-SPECIFIC FOUNDATION MODELS

Domain-specific foundation models market map

Market direction 3D image generation Code

Each vertical application encourages founders to train custom LLMs. Breakthroughs have
been made across a range of industries as researchers find benefits to combining transformer
architecture with quantitative data. Most of these research efforts focus on language or image
Customer support Geospatial
data for knowledge search and content generation. Unique data collection and labeling methods
enable DFMs in image-heavy domains including 3D images and retail along with language-focused
domains like customer support, code, and legal. In healthcare, DFMs exist both for language
(Hippocratic AI) and images (Hoppr). To complement these well-studied modalities, researchers
and startups are developing time series foundation models that inherently incorporate industry- Healthcare Legal

specific datasets of tabular data yet can apply across industries. Nixtla, the developer of time
series foundation model Time GPT, found a suitable application of time series models in financial
forecasting, where conventional language models have little ability to pattern match with new
financial data. Google Research recently published research on a competitive model called Materials Retail
TimesFM that also performs well across numerous unseen tabular datasets, with an initial focus
on retail demand planning. Impressive results across these domains make it unlikely that a base
model like OpenAI’s will succeed commercially in every industry.
Time series Robotics

Q4 2023 Artificial Intelligence & Machine Learning Report CONFIDENTIAL. NOT FOR REDISTRIBUTION. PG 21
DOMAIN-SPECIFIC FOUNDATION MODELS

Transformer architecture encourages aggregation and open access to unique datasets that can At the early stage, domain-specific model companies can raise large seed rounds to focus on high-
enable new foundation models. In Q4, Alphabet subsidiary DeepMind collaborated with academic value domains such as code. In Q3 2023, code model startup Poolside AI raised a $126.0 million
researchers to open-source the Open-X Embodiment dataset for robotics. The project has already startup to develop an end-to-end application development model that learns from both software
resulted in a new robotics foundation model with RT-X and a competitive open-source model Octo development and production environments. The company disclosed model development efforts
from academic researchers. IBM and NASA collaborated on a geospatial foundation model based in collaboration with datacenter provider Scaleway that should produce a working model by mid-
on NASA’s repository of Earth-satellite data. In healthcare, academic researchers leverage unique 2024. Also, in Q3 Imbue raised a $212.0 million Series B that is focused on general-purpose agents,
clinical datasets for biomedical image classification based on novel biomarker research with yet the company initially focused on code development models to differentiate from general
promising early results. Both proprietary datasets and synthetic data remain important for these purpose model labs. In more niche domains, startups are raising lower deal values to develop
new paradigms. Academic projects are likely to commercialize over time via startups. proofs of concept at a smaller scale. We believe some of those efforts may also be deserving of
large funding rounds given their more immediate commercial applications.

Trending startups
Domain-specific models produce commercial traction for startups. Many AI core startups have
not generated large customer bases outside of other AI startups and experimental developers,
limiting their revenue growth. Bucking this trend, DFM startup Harvey has scaled to $10.0 million
in revenue via partnerships with professional services firms including PwC. In Q4, the company
raised an $80.0 million Series B at a 4.2x valuation step-up in less than a year, led by AI specialist
VCs Elad Gil and Kleiner Perkins. The company is expanding from the legal domain to other
professional services. Harvey customer PwC is developing an M&A-specific model based on
internal documents from its M&A processes that can smooth document review. This business
model resembles the custom LLM service offered by OpenAI yet may produce network effects
within large professional services industries as architectural choices yield improved performance
on individual tasks.

Q4 2023 Artificial Intelligence & Machine Learning Report CONFIDENTIAL. NOT FOR REDISTRIBUTION. PG 22
DOMAIN-SPECIFIC FOUNDATION MODELS

Recent domain-specific foundation model VC deals ($M)*

Company Close date Domain Deal value ($M) Deal type Lead investor(s)

Harvey December 19 Legal (and other professional services) $80.0 Series B Elad Gil, Kleiner Perkins

HOPPR November 26 Healthcare $3.0 Late-stage VC Health2047

Imbue September 5 Code $212.0 Series B Astera Institute, Further Future Foundation, Nvidia

Archetype Ai August 30 Time series N/A Seed Venrock

Poolside AI August 24 Code $126.0 Seed Felicis, Redpoint Ventures, Xavier Niel

Hippocratic AI July 25 Healthcare $15.0 Seed Cincinnati Children's Hospital Medical Center, HonorHealth

Source: PitchBook • Geography: Global • *As of December 31, 2023

Q4 2023 Artificial Intelligence & Machine Learning Report CONFIDENTIAL. NOT FOR REDISTRIBUTION. PG 23
Local LLMs
Overview
AI lifecycle software market size estimate by deployment type ($B)*
Local LLMs bring advanced computation from the cloud to distributed workstations, including CPUs
and smartphones. They depend on small model architectures along with model-serving techniques
$50
that include optimization engines and request batching. Small models benefit from curated datasets
that include synthetic samples of reasoning, some of which are generated by other AI models. $45
This data-centric process reduces the number of parameters. To serve these small models locally,
$40
optimization engines convert model queries into python scripts and request-batching servers create
a queue that can be managed by small hardware form factors such as laptops and phones. Local LLM $35
frameworks can support multiple types of models including open-source models and API calls to
$30
closed-source models. Current frameworks focus on small instances of LLMs and fine-tuned open-
source models including Meta’s Llama, Mistral’s Mixtral, and Microsoft’s Phi. $25

$20
On-premises workloads remain an important part of AI workloads. On-premises AI software services
contributed 33.3% of AI software services spending in 2023, according to IDC, reaching $7.9 billion. $15
While public cloud is growing much faster, on-premises spending is still on pace to double by 2027.
$10
This market size remains large enough for startup opportunities, particularly with the tailwind of
generative AI. $5

$0
Startups can address the local LLM opportunity by providing model-serving frameworks. Model
2023 2026
serving addresses the need for hardware optimization in AI models. Nvidia stands out for the
On-premises/private cloud Public cloud
developer-friendliness of its CUDA software in serving models to the company’s proprietary hardware
via compiler software, encouraging adoption of its hardware in datacenters. Nvidia has achieved
Source: IDC • Geography: Global • *As of November 27, 2023
$1.0 billion annual recurring revenue (ARR) for its software, support, and services offerings for the
datacenter hardware, which is not a high percentage of hardware sales given bundling of CUDA

Q4 2023 Artificial Intelligence & Machine Learning Report CONFIDENTIAL. NOT FOR REDISTRIBUTION. PG 24
LOCAL LLMS

software with GPU chips, yet it remains a large business that other hardware vendors would like to
Indicative internal LLM application development lifecycle for midsize to
match.6 Model-serving companies have generated large developer communities if not significant
large enterprise
financial outcomes, led by Numenta’s $352.8 million valuation from 2021.

Market direction
• Closed-source model APIs
1 month
Deployment startups can benefit from partnerships with hardware manufacturers. Chip incumbent
AMD has turned to hardware optimization startups to close its deficit in software to Nvidia, acquiring
Nod.ai for hardware deployment and partnering with Lamini for training optimization. Nod.ai had
• Experiment with retrieval-augmented generation using closed-source model and
previously been valued at $64.0 million, suggesting the acquisition may have been more than $100
vector databases
million. These partnerships are not solely intended to address local deployment but also benefit 3 months
AMD’s PC processors, a market segment in which the company forms a duopoly with Intel. LG has
partnered with VC-backed Upstage to develop on-device models. Open-source project llama.cpp
optimizes models to run on Apple M2 chips, alleviating the need to buy GPUs to run LLMs. Apple • Fine tuning open-source models using data samples from vector database
previously acquired local machine learning model startup Xnor.ai for $200.0 million and may be 1 month
similarly encouraged to find advanced methods to compress internally developed open-source
models to run on its hardware.
• Train custom small language or image generation model (SLM) from vector databases
Large enterprises are encouraged to consider small language models to run locally after • Deploy to production
1 month
experimenting. We have heard from leading AI consultants that sophisticated customers opt to
train small language models and run them on self-hosted hardware once they develop sufficient
databases. 25% of large enterprises currently self-host models, a figure we believe includes the most
sophisticated users.7 According to the consultants, power users are encouraged by the cost savings

6: “Investor Presentation Q3 FY24,” Nvidia, November 27, 2023

7: “State of AI: A 2023 Report on AI in Production,” Retool, November 13, 2023.

Q4 2023 Artificial Intelligence & Machine Learning Report CONFIDENTIAL. NOT FOR REDISTRIBUTION. PG 25
LOCAL LLMS

and accuracy in specific tasks offered by Microsoft’s latest model Phi, which incorporates only 2 billion
GitHub star count for selected open-source local large language model
parameters—an infinitesimal fraction of those in leading models such as GPT-4. This technique can
repositories
benefit from local LLMs. While independent developers have been leading proponents of local models,
we believe that lengthy enterprise pilot projects are leading to custom training of local models.
55,000

50,000
Trending startups
45,000
We track hardware acceleration and edge deployment startups as a standalone category of our large 40,000
language model operations (LLMOps) market map. The fastest-growing startups come from the open-
35,000
source community and have not yet required VC funding. Open-source projects by Ollama, Llama.
cpp, and Oobabooga have escalated rapidly in developer adoption since Q2 2023. Ollama is on a 30,000

particularly hyperbolic trajectory, as it has grown in sync with new models from open-source startup 25,000
Mistral. The project’s local inference engine has become part of reference architectures from leading
20,000
orchestration framework builders including LangChain and LlamaIndex. The project has raised seed
15,000
funding from Index Ventures in the past and may be in line for a Series A.
10,000
In Q4, Armada raised a $40.0 million Series A led by blue-chip deep tech investors Founders Fund 5,000
and Lux Capital to bring LLMs to the edge. The company is developing local datacenters built inside
0
shipping containers that are intended to bring AI computing to remote regions and internet-of-things
Mar 15 May 14 Jul 13 Sep 11 Nov 10 Jan 9
(IoT) applications. The team is led by former DataRobot executive Dan Wright, joined by big tech
2023 2024*
alums. The company directs its efforts to defense and industrial applications rather than large LLM
Ollama Oobabooga Jan Llama.cpp Local Pilot Google Cloud Local LLM PrivateGPT
use cases, making the company relatively underexposed to the generative AI wave. Early-stage
startup Arcee.ai raised a $5.5 million round in Q1 2024 to adapt small language models (SLMs) for
Source: GitHub Star History • Geography: Global • *As of February 12, 2024
private cloud usage. The startup was founded by former Hugging Face engineers who have unique
insights into the open-source model ecosystem.

Q4 2023 Artificial Intelligence & Machine Learning Report CONFIDENTIAL. NOT FOR REDISTRIBUTION. PG 26
Select company highlights
SELECT COMPANY HIGHLIGHTS: ANTHROPIC

Key company information

Founded Total raised Lead investor(s)

2021 $7.3B Alphabet, Amazon, LG CNS, Calm
Overview Ventures, SK Telecom, Salesforce
Employees Last financing valuation Ventures, Zoom Ventures, Spark Capital,
300 in 8 offices globally $25.0B Sam Bankman-Fried, Jaan Tallinn
Anthropic was founded in 2021 by a group of seven former senior members of OpenAI. Several
of the founders worked at Google Research before OpenAI. The founders split off from OpenAI Last financing
with a desire to train safe AI systems that could mitigate the worst risks of superintelligence, Raised $300.0M in a Series G
including human extinction. To serve this mission, the company was established as a public-
benefit corporation. Since its founding, Anthropic has developed proprietary AI systems and
commercialized a series of large language models called Claude.

Claude leverages transformer architecture and a reinforcement learning process referred to as

Constitutional AI. Constitutional AI tests the alignment of model outputs against a list of ethical Exit Predictor
rules and principles. In this way, the model self-improves by selecting outputs that conform to
the company’s ethics. The Constitutional AI process improves the safety of responses without
extensive human intervention such as OpenAI uses via its outsourced reinforcement learning from
human feedback approach. We have not tracked any patent for this technique. Because of this
process, Anthropic does not shift its response types as often as ChatGPT, which faces complaints
of censorship and model drift as the model changes over time. IPO: 93% probability M&A: 4% probability No exit: 3% probability

Note: Probability data based on PitchBook VC Exit Predictor Methodology.

Q4 2023 Artificial Intelligence & Machine Learning Report CONFIDENTIAL. NOT FOR REDISTRIBUTION. PG 28
SELECT COMPANY HIGHLIGHTS: ANTHROPIC

Leadership a managed Claude service on its Bedrock deployment platform. This partnership is a superpower
for Anthropic given AWS’s leading position in the cloud service provider market. Amazon has
The company is co-led by CEO & Co-founder Dario Amodei and President and Co-founder Daniela been leading in generative AI hiring while announcing new services at its Re:Invent conference,
Amodei. The two are siblings and neither reports to the other. At OpenAI, Dario Amodei focused demonstrating the investment being made into the field.
on research and Daniela Amodei on safety and policy. Dario Amodei formerly achieved a PhD in
biophysics from Princeton and worked as a research scientist at Baidu and Google. Daniela Amodei Anthropic will be a primary beneficiary of existing AWS customers leveraging the hyperscaler’s
comes from a non-technical background and formerly was a risk manager at Stripe and led human Bedrock LLM service. Additionally, significant AWS consulting relationships from Accenture,
resources and business operations at OpenAI along with technical safety teams. Deloitte, and Slalom will encourage the use of Anthropic models. We have already seen Accenture
highlight Anthropic in an AWS reference architecture.8 Accenture is the leading AI services firm
Other key co-founder executives include Head of Compute Engineering Tom Brown and Head of by market share. Given that consultants present the first port of call for enterprise deployments,
Policy Jack Clark. Brown led the engineering of OpenAI’s GPT-3 and was formerly a startup CTO and Anthropic can benefit from close relationships via AWS.
Google researcher. Jack Clark was formerly a tech journalist and led communications at OpenAI.

Outlook
Competitors
Anthropic presents a legitimate alternative to OpenAI given the company’s research talent,
Anthropic competes in the foundation model category of AI against OpenAI, Microsoft, Google, partnership with AWS, and focus on safety issues that will be a critical concern for enterprise
Meta, and the open-source community. OpenAI, Meta, and Google are developing comparable customers. The company disclosed reaching $200.0 million in ARR in 2023. Enterprise accounts
LLMs. Recently, open-source foundation model startup Mistral matched Anthropic’s performance drive this financial total, with anchor customers including Notion, Quora, and DuckDuckGo. The
despite starting two years later with fewer resources. Google has also made rapid progress with its company forecasts 325% growth in 2024 to $850.0 million ARR, demonstrating the traction
Gemini Ultra model. OpenAI’s GPT-5 is on pace to set a new bar for performance this year, which the company’s models have in enterprises and the potential for Anthropic to become a larger
will require Claude to keep pace via successive model releases. enterprise vendor than OpenAI. Successive financing rounds are likely to occur at a steep increase
given recent discussions about a $18.4 billion valuation. Current investment opportunities include
Anthropic partners with AWS and Google via cloud investments. The company raised a $4.0 billion secondary shares from prior transactions along with a special-purpose vehicle (SPV) from
investment led by Amazon that involved both cloud credits and equity investment. A month later, Menlo Ventures.
the company raised a $2.0 billion round led by Google. Based on these partnerships, AWS offers
8: “Accenture Creates a Knowledge Assist Solution Using Generative AI Services on AWS,” AWS, October 5, 2023.

Q4 2023 Artificial Intelligence & Machine Learning Report CONFIDENTIAL. NOT FOR REDISTRIBUTION. PG 29
SELECT COMPANY HIGHLIGHTS: ANTHROPIC

Financing history

Series A Series B Series C Late-stage VC Late-stage VC Late-stage VC

May 28, 2021 April 29, 2022 May 23, 2023 August 13, 2023 September 25, 2023 October 27, 2023

Total raised Total raised Total raised Total raised Total raised Total raised
$124.0M $580.0M $450.0M $100.0M $4.0B $2.0B
Pre-money valuation Pre-money valuation Pre-money valuation Pre-money valuation Pre-money valuation Pre-money valuation
$550.0M $3.4B $4.6B N/A N/A $5.8B
Lead investor(s) Lead investor(s) Lead investor(s) Lead investor(s) Lead investor(s) Lead investor(s)
Jaan Tallinn Sam Bankman-Fried Salesforce Ventures, LG CNS, Calm Ventures and Amazon Andreessen Horowitz
Zoom Ventures, Spark Capital SK Telecom

Series G

February 1, 2021

Total raised
$1.0B
Pre-money valuation
$25.0B (estimated)
Lead investor(s)
Alphabet

Q4 2023 Artificial Intelligence & Machine Learning Report CONFIDENTIAL. NOT FOR REDISTRIBUTION. PG 30
SELECT COMPANY HIGHLIGHTS: DATABRICKS

Key company information

Founded Total raised Lead investor(s)

2013 $4.2B Nvidia, T. Rowe Price, Counterpoint
Overview Global, Franklin Templeton Investments,
Employees Last financing valuation Andreessen Horowitz, NEA
7,939 in 25 offices globally $43.2B
Founded in 2013, Databricks offers a data science platform that includes AI-as-a-service (AIaaS)
functionality with a suite of data science tools for data engineering, data warehousing, and ML December 2023 Last financing
algorithms. The company grew out of the open-source Apache Spark data science community Leader in Gartner Magic Raised $684.6M in a Series I
Quadrant for Cloud Database
and created an extensible product, called the Unified Data Analytics Platform, that can ingest data
Management Systems
from enterprise silos and prepare it for cluster-based computing. Once an effective open-source
product was in place, the company moved to a closed-source model and rapidly increased revenue
beginning in 2016.

Databricks now offers a product suite on top of Apache Spark that includes an AI platform called Exit Predictor
Lakehouse AI, which is optimized for generative AI with data storage, a model training pipeline,
and model-serving and monitoring; a data lake, which integrates unstructured data in a central
database for analytics; and data security. The company has focused on the GenAI opportunity,
training a custom LLM called Dolly that can run more efficiently than GPT-4 and achieve
competitive results on question-answering tasks. Databricks recently acquired Einblick for data
analysis using generative AI. IPO: 91% probability M&A: 7% probability No exit: 2% probability

Note: Probability data based on PitchBook VC Exit Predictor Methodology.

Q4 2023 Artificial Intelligence & Machine Learning Report CONFIDENTIAL. NOT FOR REDISTRIBUTION. PG 31
SELECT COMPANY HIGHLIGHTS: DATABRICKS

Leadership Outlook
Databricks’ management has high technical expertise capable of generating product-led growth Databricks has achieved the necessary scale and growth to realize high performance in public
and has increased its public-company experience. The company was founded by seven early markets. The company neared $2.0 billion in revenue in 2023 with 50% revenue growth.9 The
contributors to Apache Spark, and all the co-founders remain at the company. CEO and Co-founder company will need to defend itself against ML-optimized challengers that can carry out more
Ali Ghodsi earned a PhD in distributed computing and was vice president of engineering before efficient streaming data analysis and in-memory analytics within its database. Furthermore,
being promoted to CEO at the start of the company’s growth phase in 2016. The acquisition of it will benefit from analytics built on its database becoming essential to high-value industries,
MosaicML brought in the Vice President of Generative AI, Naveen Rao, who has credibility in the including IT and industrial. The company remains innovative and active in future-proofing its
open-source AI community. The board features co-founders and VC investors, including 16z Co- business, leading us to believe that, in the long run, Databricks will be valued more highly than
founder Ben Horowitz, yet still lacks the public company leadership we believe is necessary to relational database incumbents based on the growth opportunities of data lake architecture for
go public. streaming data.

Competitors
Databricks stands out in the artificial intelligence & machine learning (AI & ML) market
by supporting customized AI model training using open-source models atop its data lake,
functionality that distinguishes it from legacy vendors such as SAP, Microsoft, and AWS while also
being more AI-oriented than next-generation data-warehouse vendor Snowflake. Because of its
ability to support batch processing for ML analytics and other high-value workloads, Databricks is
capturing market share from incumbents such as Alteryx, IBM, Oracle, and SAS. Databricks runs
on Microsoft Azure and AWS, although those hyperscalers would likely prefer to natively support
their GenAI workflows. The company gains market credit for data governance via its Unity Catalog
product—a critical problem for data scientists new to AI.

9: “Databricks revenue, growth, and valuation,” Sacra, February 2, 2024.

Q4 2023 Artificial Intelligence & Machine Learning Report CONFIDENTIAL. NOT FOR REDISTRIBUTION. PG 32
SELECT COMPANY HIGHLIGHTS: DATABRICKS

Financing history

Series A Series B Series C Series D Series E Series F

September 24, 2013 December 31, 2014 December 15, 2016 August 22, 2017 February 5, 2019 October 22, 2019

Total raised Total raised Total raised Total raised Total raised Total raised
$14.0M $33.0M $60.0M $140.0M $250.0M $400.0M
Pre-money valuation Pre-money valuation Pre-money valuation Pre-money valuation Pre-money valuation Pre-money valuation
$34.0M $214.0M $453.0M $800.0M $2.5B $5.8B
Lead investor(s) Lead investor(s) Lead investor(s) Lead investor(s) Lead investor(s) Lead investor(s)
Andreessen Horowitz NEA NEA Andreessen Horowitz Andreessen Horowitz Andreessen Horowitz

Series G Series H Series I

February 1, 2021 Date undisclosed Date undisclosed

Total raised Total raised Total raised

$1.0B $1.6B $684.6M
Pre-money valuation Pre-money valuation Pre-money valuation
$27.0B $36.4B $4.2B
Lead investor(s) Lead investor(s) Lead investor(s)
Franklin Templeton Investments Counterpoint Global Nvidia, T. Rowe Price

Q4 2023 Artificial Intelligence & Machine Learning Report CONFIDENTIAL. NOT FOR REDISTRIBUTION. PG 33
 PitchBook Data, Inc.
John Gabbert Founder, CEO

Nizar Tarhuni Vice President, Institutional Research and Editorial

Paul Condra Head of Emerging Technology Research

About PitchBook Additional research

Eric Bellomo
[email protected]
Derek Hernandez
[email protected]

Industry and Technology Gaming

E-Commerce
Enterprise SaaS
Infrastructure SaaS

Research
Brendan Burke Ali Javaheri
[email protected] [email protected]
Internet of Things Emerging Spaces
Information Security
Artificial Intelligence & Machine Robert Le
Learning [email protected]
Independent, objective, and timely market intel Insurtech
Aaron DeGagne Crypto
As the private markets continue to grow in complexity and competition, it’s essential for [email protected]
investors to understand the industries, sectors, and companies driving the asset class. Medtech John MacDonagh
Digital Health [email protected]
Carbon & Emissions Tech
Our Industry and Technology Research provides detailed analysis of nascent tech sectors so Alex Frederick Clean Energy Tech
[email protected]
you can better navigate the changing markets you operate in—and pursue new opportunities
Agtech Rebecca Springer
with confidence. Foodtech [email protected]
Healthcare Services
Jonathan Geurkink Healthcare IT
[email protected]
©2024 by PitchBook Data, Inc. All rights reserved. No part of this publication may be reproduced in any form or by any Supply Chain Tech Rudy Yang
means—graphic, electronic, or mechanical, including photocopying, recording, taping, and information storage and Mobility Tech [email protected]
retrieval systems—without the express written permission of PitchBook Data, Inc. Contents are based on information Enterprise Fintech
from sources believed to be reliable, but accuracy and completeness cannot be guaranteed. Nothing herein should be Kazi Helal Retail Fintech
construed as any past, current or future recommendation to buy or sell any security or an offer to sell, or a solicitation of [email protected]
an offer to buy any security. This material does not purport to contain all of the information that a prospective investor Biopharma
may wish to consider and is not to be relied upon as such or used in substitution for the exercise of independent judgment. Pharmatech